Image enhancement model training method, imaging method, scanning equipment and medium

Abstract

An image enhancement model training method, an imaging method, scanning equipment and a medium are applied to positron emission tomography equipment, and comprise the following steps: inputting a plurality of randomly generated motif images into a motif simulation system for simulation to obtain motif simulation data; dividing the motif simulation data into data with interaction depth information and data without interaction depth information according to whether spatial positions of ray and crystal action in the motif simulation data are layered or not; and performing image reconstruction on the data to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information, constructing the reconstructed data into a training data set, and performing training by using the training data set to obtain an image enhancement model. The reconstructed image with equivalent interaction depth information is obtained according to the image enhancement model, visual errors of the interaction depth effect are reduced under the condition that the structure of hardware equipment is not changed, and the resolution and uniformity of the image are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to an image enhancement model training method, an imaging method, a scanning device, and a medium. Background Art

[0002] Positron Emission Tomography (PET) is a nuclear medicine imaging technology that generates images by detecting the annihilation process of positrons and electrons emitted by radioactive tracers, and can be used to observe and evaluate biological processes and disease states in living organisms. It has very wide applications in the fields of cancer, heart disease, and neurology.

[0003] While large-bore whole-body PET systems are widely used in clinical diagnosis, small-field-of-view PET devices dedicated to specific organs have gradually become a research hotspot, such as brain PET, small-animal PET, and breast PET. PET devices dedicated to local organs have the advantages of high sensitivity, light weight, and low cost. Their aperture is generally only one-tenth to one-half of that of the whole-body system. However, the smaller system aperture significantly increases the possibility that 511 keV photons are detected at an oblique angle when the rays interact with adjacent crystal bars, thus triggering a parallax error called the Depth of Interaction (DOI) effect. This effect is particularly obvious in the edge region, and the inconsistency of the response lines is more serious than that in the central region, seriously affecting the spatial resolution and spatial resolution uniformity of the reconstructed image. Therefore, it is necessary to reduce the parallax error in the PET reconstructed image and improve the spatial resolution non-uniformity. Summary of the Invention

[0004] The main technical problem to be solved by the present invention is that there are visual errors in the PET reconstructed image, and the spatial resolution of the image is non-uniform. Moreover, improving the image quality by changing the structure of the hardware device will lead to a significant increase in system cost and complexity, and it is difficult to maintain.

[0005] According to a first aspect, in one embodiment, an image enhancement model training method is provided, which is applied to a positron emission tomography device and includes:

[0006] Obtain multiple randomly generated phantom images, input the multiple phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data; the multiple phantom images are randomly generated according to a preset heat source rod; the phantom simulation data includes the spatial positions of the interaction between rays and crystals;

[0007] Divide the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the interaction between the rays and the crystal in the phantom simulation data are stratified, and perform image reconstruction on the data with interaction depth information and the data without interaction depth information respectively to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information;

[0008] Construct the reconstructed image with interaction depth information and the reconstructed image without interaction depth information into a training data set, and train an image enhancement model according to the training data set.

[0009] In one embodiment, the obtaining a plurality of randomly generated phantom images includes:

[0010] Obtain a plurality of heat source rods with different diameters, and construct a plurality of triangular sectors according to the plurality of heat source rods; wherein, the center distance between any two adjacent heat source rods in the triangular sector is twice the diameter of the heat source rod;

[0011] Perform a combination process on the plurality of triangular sectors to obtain a plurality of phantom images.

[0012] In one embodiment, the dividing the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the interaction between the rays and the crystal in the phantom simulation data are stratified includes:

[0013] When the spatial positions of the interaction between the rays and the crystal in the phantom simulation data are not stratified, divide the phantom simulation data into data without interaction depth information;

[0014] When the spatial positions of the interaction between the rays and the crystal in the phantom simulation data are stratified, divide the phantom simulation data into data with interaction depth information.

[0015] In one embodiment, the position coordinates of the spatial position of the interaction between the rays and the crystal in the data without interaction depth information are located at the center of the crystal;

[0016] The position coordinates of the spatial position of the interaction between the rays and the crystal in the data with interaction depth information are located at the center of a certain layer after stratifying along the crystal depth.

[0017] In one embodiment, before the step of inputting the plurality of phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data, the method further includes:

[0018] Construct a phantom simulation system using a preset simulation tool, and set the acquisition time, source activity, and physical process of the phantom simulation system.

[0019] In one embodiment, the image enhancement model trained according to the training dataset includes:

[0020] Use the reconstructed images with interaction depth information in the training dataset as input label data, and use the reconstructed images without interaction depth information in the training dataset as input image data. Construct corresponding image label data pairs according to the input label data and the input image data;

[0021] Divide the image label data pairs into a training set, a validation set, and a test set according to a preset division ratio;

[0022] Use a predefined data augmentation method to perform image transformation processing on the training set pairs to obtain augmented data pairs;

[0023] Use the augmented data pairs to train a preset augmentation model until a preset model convergence judgment condition is met to obtain a trained augmentation model; wherein, the preset augmentation model consists of a generator and a discriminator;

[0024] Use the validation set to verify the trained augmentation model. When a preset verification condition is met, obtain a verified augmentation model;

[0025] Use the test set and a preset index calculation formula to perform test processing on the verified augmentation model, and use the model that meets the preset test requirements as the image enhancement model.

[0026] In one embodiment, the using the augmented data pairs to train a preset augmentation model until a preset model convergence judgment condition is met to obtain a trained augmentation model includes:

[0027] Input the augmented data pairs into the generator in the preset augmentation model, and use the feature encoder in the generator to perform encoding processing on the augmented data pairs to obtain encoded images; the feature encoder consists of a preset first number of encoding convolutional modules;

[0028] Use the feature decoder in the generator to perform image decoding on the encoded images to obtain initial generated images; the feature decoder consists of a preset second number of decoding convolutional modules;

[0029] Use multiple convolutional modules in the discriminator of the augmentation model to perform convolutional processing on the initial generated images to obtain convolutional images;

[0030] Input the convolutional image into a preset activation function to obtain a discrimination probability, and generate an image discrimination result according to the discrimination probability;

[0031] When the image discrimination result meets the preset model convergence judgment condition, use the enhancement model as the trained enhancement model.

[0032] According to a second aspect, an imaging method for a positron emission tomography device is provided in an embodiment, including:

[0033] Obtain an image reconstructed using detection data without interaction depth information;

[0034] Input the image reconstructed from the detection data without interaction depth information into a trained image enhancement model to obtain a reconstructed image with equivalent interaction depth information; the image enhancement model is obtained by training through an image enhancement model training method.

[0035] According to a third aspect, a positron emission tomography device is provided in an embodiment, including:

[0036] An image input component for obtaining an image reconstructed using detection data without interaction depth information;

[0037] An image generation component for inputting the image reconstructed from the detection data without interaction depth information into a trained image enhancement model to obtain a reconstructed image with equivalent interaction depth information; the image enhancement model is obtained by training through an image enhancement model training method.

[0038] According to the image enhancement model training method, device and medium of the above embodiments, it includes: obtaining multiple randomly generated phantom images, inputting the multiple phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data. Using the pre-constructed phantom simulation system to generate a large amount of rich phantom simulation data as training data for subsequent model training. Divide the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial position of the interaction between the ray and the crystal in the phantom simulation data is stratified, and perform image reconstruction on the divided data respectively to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information. Dividing according to the spatial position of the interaction between the ray and the crystal in the phantom simulation data can identify whether there is interaction depth information in the phantom simulation data, and train an image enhancement model based on the reconstructed image. Therefore, the trained image enhancement model can be applied to a positron emission tomography device to generate a reconstructed image with equivalent interaction depth information. Description of the Drawings

[0039] Figure 1 It is a flowchart for training an image enhancement model according to an embodiment of the present application;

[0040] Figure 2 It is a schematic diagram of the training process of an image enhancement model according to an embodiment;

[0041] Figure 3 It is a schematic diagram of the training process of an image enhancement model according to another embodiment;

[0042] Figure 4 It is a schematic diagram of the training process of an image enhancement model according to another embodiment;

[0043] Figure 5 It is a schematic diagram of the training process of an image enhancement model according to another embodiment;

[0044] Figure 6 It is a schematic diagram of the imaging process of a positron emission tomography device according to an embodiment;

[0045] Figure 7 It is a schematic diagram of the structure of a positron emission tomography device according to an embodiment. Detailed implementation manners

[0046] The present invention will be further described in detail below in conjunction with the accompanying drawings through specific implementation manners. Similar elements in different implementation manners are labeled with related similar element numbers. In the following implementation manners, many details are described to enable a better understanding of the present application. However, those skilled in the art can easily recognize that some of the features can be omitted in different situations, or can be replaced by other elements, materials, or methods. In some cases, some operations related to the present application are not shown or described in the specification to avoid the core part of the present application being overwhelmed by excessive description. For those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0047] In addition, the features, operations, or characteristics described in the specification can be combined in any appropriate manner to form various implementation manners. At the same time, the steps or actions in the method description can also be reordered or adjusted in a manner obvious to those skilled in the art. Therefore, the various sequences in the specification and drawings are only for clearly describing a certain embodiment and do not mean a necessary sequence unless it is stated that a certain sequence must be followed.

[0048] The serial numbers assigned to the components in this text itself, such as "first", "second", etc., are only used to distinguish the described objects and do not have any sequential or technical meanings. And the "connection" and "coupling" mentioned in this application, unless otherwise specified, both include direct and indirect connection (coupling).

[0049] In some embodiments, multiple randomly generated phantom images are obtained, and the multiple phantom images are input into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data. A large amount of rich phantom simulation data is generated by using the pre-constructed phantom simulation system as training data for subsequent model training. The phantom simulation data is divided into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the interaction between the rays and the crystals in the phantom simulation data are stratified, and the divided data is respectively subjected to image reconstruction to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information. By dividing according to the spatial positions of the interaction between the rays and the crystals in the phantom simulation data, it is possible to identify whether there is interaction depth information in the phantom simulation data, and an image enhancement model is trained based on the reconstructed images. Therefore, the trained image enhancement model can be applied to a positron emission tomography device to generate a reconstructed image with equivalent interaction depth information.

[0050] Please refer to Figure 1 , some embodiments of the present invention provide a method for training an image enhancement model, including steps S10 to step S30, which will be specifically described below.

[0051] Step S10: Obtain multiple randomly generated phantom images, and input the multiple phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data; the multiple phantom images are randomly generated according to a preset heat source rod; the phantom simulation data includes the spatial positions of the interaction between the rays and the crystals.

[0052] Please refer to Figure 2 , in some embodiments, step S10 of obtaining multiple randomly generated phantom images includes steps S11 to step S12, which will be specifically described below.

[0053] Step S11: Obtain multiple heat source rods with different diameters, and construct multiple triangular sectors according to the multiple heat source rods; wherein, the center distance between any two adjacent heat source rods in the triangular sector is twice the diameter of the heat source rod.

[0054] In some embodiments, the different diameters of the multiple heat source rods with different diameters usually refer to being randomly generated in the diameter range of 1.8 - 5 mm. Corresponding triangular sectors can be constructed according to the multiple heat source rods. Among them, the center distance between any two adjacent heat source rods in the same triangular sector is twice the diameter of the heat source rod.

[0055] Step S12: Combine multiple triangular sectors to obtain multiple phantom images.

[0056] In some embodiments, each phantom image contains six triangular sectors. Therefore, combine the six triangular sectors into the corresponding phantom image. Among them, the phantoms in each phantom image are randomly arranged in the imaging range (FOV, Field of View).

[0057] In some embodiments, before inputting multiple phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data, it further includes:

[0058] Use a preset simulation tool to construct a phantom simulation system, and set the acquisition time, source activity, and physical process of the phantom simulation system.

[0059] In some embodiments, the preset simulation tool is the Gate simulation tool. Use the Gate simulation tool to build a PET system dedicated to the brain, simulate the actual acquisition scenario, and the source activity of the phantom simulation system can be set to 10 kBq / mL, and the acquisition time is 5 min.

[0060] Step S20: Divide the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the interaction between the rays and the crystals in the phantom simulation data are stratified, and perform image reconstruction on the data with interaction depth information and the data without interaction depth information respectively to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information.

[0061] Please refer to Figure 3 , in some embodiments, step S20 divides the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the interaction between the rays and the crystals in the phantom simulation data are stratified, including steps S21 to S22, which are specifically described below.

[0062] Step S21: When the spatial positions of the interaction between the rays and the crystals in the phantom simulation data are not stratified, divide the phantom simulation data into data without interaction depth information.

[0063] Step S22: When the spatial positions of the interaction between the rays and the crystals in the phantom simulation data are stratified, divide the phantom simulation data into data with interaction depth information.

[0064] In some embodiments, when the spatial positions of the interactions between the rays and the crystals in the phantom simulation data are not stratified, the position coordinates of the spatial positions of the interactions between the rays and the crystals are located at the center of the crystal. When the spatial positions of the interactions between the rays and the crystals in the phantom simulation data are stratified, the position coordinates of the spatial positions of the interactions between the rays and the crystals are located at the center of a certain layer after stratification along the depth of the crystal. Assuming that the spatial positions of the interactions between the rays and the crystals in the phantom simulation data are divided into four layers, then the position coordinates of the spatial positions include four possibilities and can be located at the center of one of the four different corresponding layers.

[0065] In some embodiments, since no actual system can measure the actual position of the ray interaction and can only achieve a certain accuracy, then this accuracy determines the number of layers when the crystal is stratified.

[0066] In some embodiments, image reconstruction is performed separately on the data with interaction depth information and the data without interaction depth information, and a reconstructed image with interaction depth information and a reconstructed image without interaction depth information can be obtained. Among them, performing image reconstruction separately on the data with interaction depth information and the data without interaction depth information is achieved by using the OP-OSEM (Ordinary-Poisson Ordered-Subset Expectation-Maximum) iterative algorithm.

[0067] Step S30: Construct a training data set from the reconstructed image with interaction depth information and the reconstructed image without interaction depth information, and train an image enhancement model according to the training data set.

[0068] Reference Figure 4 , in some embodiments, step S30 of training an image enhancement model according to the training data set includes steps S31 to S36, which are specifically described below.

[0069] Step S31: Use the reconstructed image with interaction depth information in the training data set as input label data, use the reconstructed image without interaction depth information in the training data set as input image data, and construct corresponding image label data pairs according to the input label data and the input image data.

[0070] Step S32: Divide the image label data pairs into a training set, a validation set, and a test set according to a preset division ratio.

[0071] Step S33: Perform image transformation processing on the training set using a predefined data enhancement method to obtain enhanced data pairs.

[0072] In some embodiments, a predefined data augmentation method is used to perform image transformation processing on the training set to increase the robustness of the model. Among them, the data augmentation processing includes rotation, mirroring, randomly adding Gaussian noise, and randomly adding Gaussian blur in random spatial transformation, etc.

[0073] In some embodiments, when performing image transformation processing on the training set, assuming there are 100 groups of image-label data pairs, different transformations will be performed on each group of image-label data pairs. For example, a rotation transformation will be performed on data pair A, but a transformation combining mirroring and randomly adding Gaussian noise will be performed on data pair B. Among them, for the same pair of image and label, only the same transformation can be used. That is, assuming the image has been rotated, the corresponding label of the image also needs to be rotated by the same transformation.

[0074] Step S34: Use the augmented data pairs to train the preset augmented model until the preset model convergence judgment condition is met, and obtain the trained augmented model; among them, the preset augmented model is composed of a generator and a discriminator.

[0075] In some embodiments, the preset augmented model includes a generator and a discriminator. During the training process, the generator and the discriminator will be alternately trained in sequence, and the training frequencies of the generator and the discriminator are controlled to be 1:1.

[0076] Reference Figure 5 , in some embodiments, step S34 uses the augmented data pairs to train the preset augmented model until the preset model convergence judgment condition is met, and obtaining the trained augmented model includes steps S341 to S345, which are specifically described below.

[0077] Step S341: Input the augmented data pairs into the generator in the preset augmented model, and use the feature encoder in the generator to perform encoding processing on the augmented data pairs to obtain encoded images; the feature encoder is composed of a preset first number of encoding convolutional modules.

[0078] In some embodiments, an improved UNet network is constructed as the generator for generating a reconstructed image with interaction depth information from a reconstructed image without interaction depth information. Among them, the improved UNet network includes a feature encoder and a feature decoder. The feature encoder is composed of four sequentially connected stacked encoding convolutional models, and the feature decoder is composed of three sequentially connected stacked decoding convolutional modules.

[0079] In some embodiments, the preset first number is four.

[0080] Step S342: Use the feature decoder in the generator to decode the encoded image to obtain an initial generated image; the feature decoder is composed of a preset second number of decoding convolutional modules.

[0081] In some embodiments, the preset second number is three.

[0082] In some embodiments, the generator is trained using a cross-entropy loss function, and the cross-entropy loss function can be expressed as where Loss G represents the loss function corresponding to the generator, G represents the generator, D represents the discriminator, and cGAN is a preset enhancement model. can be expressed as where x and y respectively represent the image and its corresponding label in the enhanced data pair, z represents random Gaussian noise, and G(x, z) represents the image generated by the generator. can be expressed as λ is a hyperparameter used to adjust and the weights of Loss G When only is used (i.e., λ = 0), clearer results can be generated, but sometimes artifacts will be introduced. After adding (λ = 100), the artifacts can be reduced.

[0083] Step S343: Use multiple convolutional modules in the discriminator of the enhancement model to perform convolutional processing on the initial generated image to obtain a convolutional image.

[0084] In some embodiments, the discriminator is a PatchGAN network used to discriminate whether an image is a real reconstructed image, and the network is sequentially stacked by a series of convolutional modules including convolutional layers.

[0085] In some embodiments, the discriminator uses a least squares loss function, and the least squares loss function is where Loss D is the loss function corresponding to the discriminator, G represents the generator, D represents the discriminator, and cGAN is a preset enhancement model. can be expressed as where x and y respectively represent the image and its corresponding label in the enhanced data pair, z represents random Gaussian noise, and G(x, z) represents the image generated by the generator.

[0086] Step S344: Input the convolutional image into a preset activation function to obtain a discrimination probability, and generate an image discrimination result according to the discrimination probability.

[0087] In some embodiments, the discrimination probability ranges from [0, 1], and an image discrimination result is generated based on the discrimination probability and a preset discrimination condition. When calculating the discriminator Loss D function, a larger image discrimination result (probability close to 1) will increase the Loss D , and during the stage of updating the discriminator weight parameters, the amplitude of the updated parameters will be increased. On the contrary, when calculating the discriminator Loss G function, a smaller image discrimination result (probability close to 0) will increase the Loss G , and during the stage of updating the generator parameters, the amplitude of the updated parameters will be increased.

[0088] Step S345: When the image discrimination result meets the preset model convergence judgment condition, the enhancement model is used as the trained enhancement model.

[0089] Step S35: Use the validation set to perform model validation on the trained enhancement model, and obtain the enhancement model that passes the validation when the preset validation conditions are met.

[0090] Step S36: Use the test set and the preset index calculation formula to perform test processing on the enhancement model that passes the validation, and use the model that meets the preset test requirements as the image enhancement model.

[0091] In some embodiments, the preset index calculation formula can be the formula of SSIM (Structural Similarity), or the formula of NRMSE (Normalized root mean square error), or the formula of PSNR (Peak Signal to Noise Ratio). Through the test processing, the recovery function of the model for images can be judged.

[0092] Reference Figure 6 , in some embodiments, an imaging method for a positron emission tomography device includes steps S40 to S50, which will be specifically described below.

[0093] Step S40: Obtain an image reconstructed using detection data without interaction depth information.

[0094] In some embodiments, the image reconstructed using detection data without interaction depth information is collected on an actual brain system.

[0095] Step S50: Input the reconstructed image of the detection data without interaction depth information into the trained image enhancement model to obtain a reconstructed image with equivalent interaction depth information; the image enhancement model is trained by the training method in any one of Steps S10 to S30.

[0096] According to the image enhancement model training method, device and medium of the above embodiment, it includes: obtaining multiple randomly generated phantom images, inputting the multiple phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data. Using the pre-constructed phantom simulation system to generate a large amount of rich phantom simulation data as the training data during subsequent model training. Divide the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the ray and crystal interactions in the phantom simulation data are stratified, and perform image reconstruction on the divided data respectively to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information. Dividing according to whether the spatial positions of the ray and crystal interactions in the phantom simulation data are stratified can identify whether there is interaction depth information in the phantom simulation data, and train an image enhancement model based on the reconstructed images. Therefore, the trained image enhancement model can be applied in a positron emission tomography device to generate a reconstructed image with equivalent interaction depth information.

[0097] Reference Figure 7 , in some embodiments, a positron emission tomography device is provided, including an image input component 10 and an image generation component 20, which will be specifically described below.

[0098] The image input component 10 is used to obtain an image reconstructed using detection data without interaction depth information.

[0099] In some embodiments, the image reconstructed using detection data without interaction depth information is collected on an actual brain system.

[0100] The image generation component 20 is used to input the image of the detection data without interaction depth information into the trained image enhancement model to obtain a reconstructed image with equivalent interaction depth information; the image enhancement model is trained by the image enhancement model training method, and the following operations can be performed during the training of the image enhancement model, which will be specifically described below.

[0101] The image generation component 20 is used to obtain multiple randomly generated phantom images, input the multiple phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data; the multiple phantom images are randomly generated according to a preset heat source rod; the phantom simulation data includes the spatial positions of the ray and crystal interactions.

[0102] Please return to the reference Figure 2 In some embodiments, the image generation component 20 can perform the following actions when obtaining multiple randomly generated phantom images, which will be specifically described below.

[0103] The image generation component 20 obtains multiple heat source rods with different diameters, and constructs multiple triangular sectors according to the multiple heat source rods; wherein, the center distance between any two adjacent heat source rods in the triangular sector is twice the diameter of the heat source rod, and the multiple triangular sectors are combined to obtain multiple phantom images.

[0104] In some embodiments, the different diameters of the multiple heat source rods with different diameters are usually randomly generated within the diameter range of 1.8 - 5 mm. Corresponding triangular sectors can be constructed according to the multiple heat source rods. Among them, the center distance between any two adjacent heat source rods in the same triangular sector is twice the diameter of the heat source rod.

[0105] In some embodiments, each phantom image contains six triangular sectors. Therefore, the six triangular sectors are combined into corresponding phantom images. Among them, the phantoms in each phantom image are randomly arranged within the imaging range (FOV, Field of View).

[0106] In some embodiments, before inputting the multiple phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data, it further includes:

[0107] Using a preset simulation tool to construct a phantom simulation system, and setting the acquisition time, source activity, and physical process of the phantom simulation system.

[0108] In some embodiments, the preset simulation tool is the Gate simulation tool. Using the Gate simulation tool to build a PET system dedicated to the brain, simulating the actual acquisition scenario, the source activity of the phantom simulation system can be set to 10 kBq / mL, and the acquisition time is 5 min.

[0109] The image generation component 20 divides the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the rays and the crystals in the phantom simulation data are stratified, and respectively performs image reconstruction on the data with interaction depth information and the data without interaction depth information to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information.

[0110] Please return to the reference Figure 3, in some embodiments, the image generation component 20 can perform the following actions according to whether the spatial positions of the interactions between the rays and the crystal in the phantom simulation data are stratified. The following is a specific description.

[0111] When the spatial positions of the interactions between the rays and the crystal in the phantom simulation data are not stratified, the phantom simulation data is divided into data without interaction depth information. When the spatial positions of the interactions between the rays and the crystal in the phantom simulation data are stratified, the phantom simulation data is divided into data with interaction depth information.

[0112] In some embodiments, when the spatial positions of the interactions between the rays and the crystal in the phantom simulation data are not stratified, the position coordinates of the spatial positions of the interactions between the rays and the crystal are located at the center of the crystal. When the spatial positions of the interactions between the rays and the crystal in the phantom simulation data are stratified, the position coordinates of the spatial positions of the interactions between the rays and the crystal are located at the center of a certain layer after stratification along the crystal depth. Assuming that the spatial positions of the interactions between the rays and the crystal in the phantom simulation data are divided into four layers, then the position coordinates of the spatial positions include four possibilities and can be located at the center of one of the four different corresponding layers.

[0113] In some embodiments, since no actual system can measure the actual position of the ray interaction and can only achieve a certain accuracy, then this accuracy determines the number of layers when the crystal is stratified.

[0114] In some embodiments, image reconstruction is respectively performed on the data with interaction depth information and the data without interaction depth information, and a reconstructed image with interaction depth information and a reconstructed image without interaction depth information can be obtained. Among them, image reconstruction is respectively performed on the data with interaction depth information and the data without interaction depth information by using the OP-OSEM (Ordinary-Poisson Ordered-Subset Expectation-Maximum) iterative algorithm.

[0115] The image generation component 20 constructs the reconstructed image with interaction depth information and the reconstructed image without interaction depth information into a training data set, and trains an image enhancement model according to the training data set.

[0116] Please return to refer to Figure 4 , in some embodiments, the image generation component 20 can perform the following actions according to the image enhancement model trained according to the training data set. The following is a specific description.

[0117] The image generation component 20 uses the reconstructed images with interaction depth information in the training dataset as input label data, and the reconstructed images without interaction depth information in the training dataset as input image data, and constructs corresponding image-label data pairs according to the input label data and input image data. The image-label data pairs are divided into a training set, a validation set, and a test set according to a preset division ratio. The training set is processed by image transformation using a predefined data augmentation method to obtain augmented data pairs. The augmented data pairs are used to train a preset augmentation model until a preset model convergence judgment condition is met, and a trained augmentation model is obtained; wherein, the preset augmentation model is composed of a generator and a discriminator. The trained augmentation model is validated using the validation set, and a validated augmentation model is obtained when a preset validation condition is met. The validated augmentation model is tested using the test set and a preset metric calculation formula, and the model that meets the preset test requirements is used as the image augmentation model.

[0118] In some embodiments, the training set is processed by image transformation using a predefined data augmentation method to increase the robustness of the model. Among them, the data augmentation process includes rotation, mirroring, randomly adding Gaussian noise, and randomly adding Gaussian blur in random spatial transformation, etc.

[0119] In some embodiments, when processing the training set by image transformation, assuming there are 100 groups of image-label data pairs, different transformations will be performed on each group of image-label data pairs. For example, the A data pair will be rotated, but the B data pair will be transformed by combining mirroring and randomly adding Gaussian noise. Among them, for the same pair of image and label, only the same transformation can be used, that is, assuming the image is rotated, then the corresponding label of the image will also be rotated by the same amount.

[0120] In some embodiments, the preset augmentation model includes a generator and a discriminator. During the training process, the generator and the discriminator will be alternately trained in turn, and the training frequencies of the generator and the discriminator are controlled to be 1:1.

[0121] Refer to the return for reference Figure 5 In some embodiments, the image generation component 20 uses the augmented data pairs to train a preset augmentation model until a preset model convergence judgment condition is met, and the trained augmentation model can perform the following actions, which will be specifically described below.

[0122] The image generation component 20 inputs the enhanced data pair into the generator in a preset enhanced model, and uses the feature encoder in the generator to encode the enhanced data pair to obtain an encoded image; the feature encoder is composed of a preset first number of encoding convolutional modules. The feature decoder in the generator is used to decode the encoded image to obtain an initial generated image; the feature decoder is composed of a preset second number of decoding convolutional modules. Multiple convolutional modules in the discriminator of the enhanced model are used to perform convolutional processing on the initial generated image to obtain a convolutional image. The convolutional image is input into a preset activation function to obtain a discrimination probability, and an image discrimination result is generated according to the discrimination probability. When the image discrimination result meets the preset model convergence judgment condition, the enhanced model is used as the trained enhanced model.

[0123] In some embodiments, an improved UNet network for generating a reconstructed image with interaction depth information from a reconstructed image without interaction depth information is constructed as the generator. Among them, a feature encoder and a feature decoder are included in the improved UNet network. The feature encoder is composed of four sequentially connected stacked encoding convolutional models, and the feature decoder is composed of three sequentially connected stacked decoding convolutional modules.

[0124] In some embodiments, the preset first number is four, and the preset second number is three.

[0125] In some embodiments, the generator is trained using a cross-entropy loss function, and the cross-entropy loss function can be expressed as where Loss G is the loss function corresponding to the generator, G represents the generator, D represents the discriminator, and cGAN is the preset enhanced model. can be expressed as where x and y respectively represent the image and its corresponding label in the enhanced data pair, z represents random Gaussian noise, and G(x,z) represents the image generated by the generator. can be expressed as λ is a hyperparameter used to adjust and the weights of Loss G When only using (i.e., λ = 0), clearer results can be generated, but sometimes artifacts are introduced. After adding (λ = 100), artifacts can be reduced.

[0126] In some embodiments, the discriminator is a PatchGAN network for discriminating whether an image is a real reconstructed image, and the network is sequentially stacked by a series of convolutional modules including convolutional layers.

[0127] In some embodiments, the discriminator adopts a least - squares loss function, and the least - squares loss function is where Loss D is the loss function corresponding to the discriminator, G represents the generator, D represents the discriminator, and cGAN is a preset enhancement model, which can be expressed as where x and y respectively represent the image and its corresponding label in the enhanced data pair, z represents random Gaussian noise, and G(x, z) represents the image generated by the generator.

[0128] In some embodiments, the range of the discrimination probability belongs to [0, 1], and the image discrimination result is generated according to the discrimination probability and the preset discrimination condition. When calculating the discriminator Loss D loss function, a larger image discrimination result (probability close to 1) of the generated image will increase Loss D , and during the stage of updating the discriminator weight parameters, the amplitude of the updated parameters will be increased. On the contrary, when calculating the discriminator Loss G loss function, a smaller image discrimination result (probability close to 0) of the generated image will increase Loss G , and during the stage of updating the generator parameters, the amplitude of the updated parameters will be increased.

[0129] In some embodiments, the preset index calculation formula can be the formula of SSIM (Structural Similarity), or the formula of NRMSE (Normalized root - mean - square error), or the formula of PSNR (Peak Signal - to - Noise Ratio). The recovery function of the model for images can be judged through test processing.

[0130] Please refer back to Figure 6 , in some embodiments, an imaging method of a positron emission tomography device can perform the following operations, which will be specifically described below.

[0131] Obtain an image reconstructed using detection data without interaction depth information, input the image reconstructed from the detection data without interaction depth information into a trained image enhancement model, and obtain a reconstructed image with interaction depth information; the image enhancement model is trained through an image enhancement model training method.

[0132] In some embodiments, the image reconstructed using detection data without interaction depth information is collected on an actual brain system.

[0133] The image enhancement model training method, device, and medium according to the above embodiments include: obtaining a plurality of randomly generated phantom images, inputting the plurality of phantom images into a pre-constructed phantom simulation system for simulation processing to obtain phantom simulation data. A large amount of rich phantom simulation data is generated by the pre-constructed phantom simulation system and used as training data during subsequent model training. The phantom simulation data is divided into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the interaction between the rays and the crystals in the phantom simulation data are stratified, and the divided data is respectively subjected to image reconstruction to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information. By dividing according to the spatial positions of the interaction between the rays and the crystals in the phantom simulation data, it is possible to identify whether there is interaction depth information in the phantom simulation data, and an image enhancement model is trained based on the reconstructed images. Therefore, the trained image enhancement model can be applied to a positron emission tomography device to generate a reconstructed image with equivalent interaction depth information.

[0134] Those skilled in the art can understand that all or part of the functions of the above methods can be implemented in a hardware manner or in a computer program manner. When all or part of the functions in the above embodiments are implemented in a computer program manner, the program can be stored in a computer-readable storage medium. The storage medium may include: read-only memory, random access memory, magnetic disk, optical disk, hard disk, etc. The above functions are realized by a computer executing the program. For example, the program is stored in the memory of the device, and when the processor executes the program in the memory, the above all or part of the functions can be realized. In addition, when all or part of the functions in the above embodiments are implemented in a computer program manner, the program can also be stored in a storage medium such as a server, another computer, magnetic disk, optical disk, flash drive, or mobile hard disk, and is saved to the memory of the local device by downloading or copying, or the system of the local device is updated. When the processor executes the program in the memory, the above all or part of the functions in the above embodiments can be realized.

[0135] The above uses specific examples to elaborate on the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art of the present invention, based on the idea of the present invention, several simple deductions, deformations, or substitutions can also be made.

Claims

1. An image enhancement model training method, applied to a positron emission tomography device, characterized in that: include: Acquire a plurality of randomly generated phantom images, and input the plurality of phantom images into a pre-built phantom simulation system for simulation processing to obtain phantom simulation data; The plurality of phantom images are randomly generated according to the preset heat source rod; the phantom simulation data includes the spatial positions of the rays and crystals; Dividing the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the rays and the crystals in the phantom simulation data are layered, and reconstructing the data with interaction depth information and the data without interaction depth information respectively to obtain a reconstructed image with interaction depth information and a reconstructed image without interaction depth information; The reconstructed image with interaction depth information and the reconstructed image without interaction depth information are constructed as a training data set, and an image enhancement model is obtained by training according to the training data set.

2. The method according to claim 1, characterized in that The step of obtaining a plurality of randomly generated motif images comprises: Acquire multiple heat source rods of different diameters, and construct multiple triangular sectors based on the multiple heat source rods; wherein the center distance between any two adjacent heat source rods in the triangular sector is twice the diameter of the heat source rod; The multiple triangular sectors are combined to obtain multiple model images.

3. The method according to claim 1, characterized in that The step of dividing the phantom simulation data into data with interaction depth information and data without interaction depth information according to whether the spatial positions of the rays and crystals in the phantom simulation data are hierarchical comprises: When the spatial positions of the rays and the crystals in the phantom simulation data are not layered, dividing the phantom simulation data into data without interaction depth information; When the spatial positions of the ray and crystal interactions in the phantom simulation data are layered, the phantom simulation data is divided into data having interaction depth information.

4. The method according to claim 1, characterized in that The position coordinates of the spatial position of the interaction between the ray and the crystal in the data without interaction depth information are located at the center of the crystal; The position coordinates of the spatial positions of the interaction between the rays and the crystal in the data with the interaction depth information are located at the center of a layer after stratification along the crystal depth.

5. The method according to claim 1, characterized in that Before inputting the plurality of phantom images into a pre-built phantom simulation system for simulation processing to obtain phantom simulation data, the method further includes: A phantom simulation system is constructed using a preset simulation tool, and the acquisition time, source activity and physical process of the phantom simulation system are set.

6. The method according to claim 1, characterized in that The step of training the image enhancement model according to the training data set comprises: Taking the reconstructed image with interaction depth information in the training data set as input label data, taking the reconstructed image without interaction depth information in the training data set as input image data, and constructing a corresponding image label data pair according to the input label data and the input image data; Dividing the image label data pairs into a training set, a validation set, and a test set according to a preset division ratio; Performing image transformation processing on the training set using a predefined data enhancement method to obtain enhanced data pairs; Using the enhanced data to train a preset enhancement model until a preset model convergence judgment condition is met, thereby obtaining a trained enhancement model; wherein the preset enhancement model is composed of a generator and a discriminator; The trained enhancement model is verified by using the verification set, and a verified enhancement model is obtained when a preset verification condition is met; The verified enhancement model is tested using the test set and the preset indicator calculation formula, and the model that meets the preset test requirements is used as the image enhancement model.

7. The method according to claim 6, characterized in that The method of using the enhanced data to perform model training on a preset enhanced model until a preset model convergence judgment condition is met to obtain a trained enhanced model includes: Inputting the enhanced data pair into a generator in a preset enhancement model, and encoding the enhanced data pair using a feature encoder in the generator to obtain an encoded image; the feature encoder is composed of a preset first number of encoding convolution modules; Using a feature decoder in the generator to decode the encoded image to obtain an initial generated image; the feature decoder is composed of a preset second number of decoding convolution modules; Using multiple convolution modules in the discriminator in the enhanced model to perform convolution processing on the initial generated image to obtain a convolution image; Input the convolution image into a preset activation function to obtain a discrimination probability, and generate an image discrimination result according to the discrimination probability; When the image discrimination result meets the preset model convergence judgment condition, the enhanced model is used as the trained enhanced model.

8. An imaging method for a positron emission tomography device, characterized in that: include: acquiring an image reconstructed using the detection data without interaction depth information; The image reconstructed from the detection data without interaction depth information is input into a trained image enhancement model to obtain a reconstructed image with equivalent interaction depth information; the image enhancement model is trained by the training method described in any one of claims 1 to 7.

9. A positron emission tomography device, characterized in that: include: An image input component for acquiring an image reconstructed using detection data without interaction depth information; An image generating component is used to input the image reconstructed by the detection data without interaction depth information into a trained image enhancement model to obtain a reconstructed image with equivalent interaction depth information; the image enhancement model is trained by the training method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The medium stores a program, which can be executed by a processor to implement the method according to any one of claims 1 to 7.

Images