Model training method and apparatus, and CT-free pet image attenuation correction method and apparatus

Abstract

The present description relates to the technical field of PET image processing, and provides a model training method and apparatus, and a CT-free PET image attenuation correction method and apparatus. The training method comprises: creating a cycle generative adversarial network comprising a first generative adversarial network and a second generative adversarial network; training the cycle generative adversarial network, so that the first generative adversarial network learns a mapping relationship that a PET image that is not subjected to attenuation correction is converted into a PET image subjected to attenuation correction, and the second generative adversarial network learns a mapping relationship that the PET image subjected to attenuation correction is converted into the PET image that is not subjected to attenuation correction; and when the cycle generative adversarial network obtained by current training satisfies a preset condition, determining a first generator in the first generative adversarial network to be a CT-free PET image attenuation correction model. The embodiments of the present description can reduce the harmful radiation dosage while ensuring the quality of PET image attenuation correction.

Description

Model training method, PET image attenuation correction method and device without CT Technical Field

[0001] This specification relates to the technical field of PET image processing, and in particular to a model training method, a PET image attenuation correction method and a device without CT. Background Art

[0002] Positron emission tomography (PET) is an advanced nuclear medicine molecular imaging technique that can visualize the distribution of radioactive tracers in a patient's body. During PET imaging, radioactive nuclides decay in the body, releasing positrons. These positrons collide with negative electrons within a range of 1-3 mm and annihilate, releasing a pair of oppositely directed gamma photons. After PET detection and reception, PET projection data is obtained. Reconstructing this projection data yields a PET image representing the distribution of metabolic levels in human tissues. However, these PET images exhibit significant noise and attenuation artifacts, resulting in low spatial resolution and unsuitable for clinical diagnosis. Therefore, attenuation correction is crucial for accurate PET quantification. In PET / CT equipment, computed tomography (CT) imaging, which provides accurate anatomical information and excellent contrast, is often used to generate attenuation correction coefficient maps for PET imaging to correct for attenuation artifacts in PET images.

[0003] However, during CT scans, patients are exposed to X-rays, which inevitably raises concerns about radiation hazards. Especially with the advent of total-body PET / CT scanners, which reduce the dose of radioactive drugs compared to whole-body PET / CT, the radiation burden of CT has become more significant. Although attenuation correction in PET imaging typically uses ultra-low-dose attenuation correction CT, low-dose CT still contributes significantly to radiation. Therefore, it is critical to research and develop CT-free attenuation correction methods for PET images, which can not only ensure the quality of attenuation correction imaging but also reduce harmful radiation doses. This has important scientific significance and application prospects in the field of medical diagnosis.

[0004] Summary of the Invention

[0005] The purpose of the embodiments of this specification is to provide a model training method, a CT-free PET image attenuation correction method and device, so as to reduce the harmful radiation dose while ensuring the quality of PET image attenuation correction.

[0006] To achieve the above objectives, on the one hand, embodiments of this specification provide a method for training a PET image attenuation correction model without CT, comprising:

[0007] Creating a cyclic generative adversarial network; the cyclic generative adversarial network includes a first generative adversarial network and a second generative adversarial network;

[0008] Training the recurrent generative adversarial network so that the first generative adversarial network learns a mapping relationship of converting an unattenuated PET image into an attenuated PET image, and the second generative adversarial network learns a mapping relationship of converting an attenuation-corrected PET image into an unattenuated PET image;

[0009] Determine whether the cyclic generative adversarial network obtained by the current training meets the preset conditions;

[0010] When the cyclic generative adversarial network obtained by the current training meets the preset conditions, the first generator in the first generative adversarial network is determined as a PET image attenuation correction model without CT.

[0011] In the PET image attenuation correction model training method of the embodiment of this specification, determining whether the currently trained cyclic generative adversarial network meets the preset conditions includes:

[0012] According to the formula L=L adv1 +L adv2 +λ1L cyc +λ2L ide +λ3L site1 +λ4L site2 Calculate the total loss value of the current training cyclic generative adversarial network;

[0013] Determining whether the total loss value is less than a loss threshold;

[0014] When the total loss value is less than the loss threshold, confirming that the cyclic generative adversarial network obtained by the current training meets the preset conditions; otherwise, confirming that the cyclic generative adversarial network obtained by the current training does not meet the preset conditions;

[0015] Among them, L is the total loss value, L adv1 is the adversarial loss value of the first generator in the first generative adversarial network, L adv2 is the adversarial loss value of the second generator in the second generative adversarial network, L cyc is the cycle consistency loss value of the first generator and the second generator, L ide is the fidelity loss value of the first generator and the second generator, L site1 is the part structure loss of the first discriminator in the first generative adversarial network, L site2is the part structure loss of the second discriminator in the second generative adversarial network, λ1, λ2, λ3 and λ4 correspond to L cyc , L ide , L site1 and L site2 The coefficient of .

[0016] In the PET image attenuation correction model training method of the embodiment of this specification, the adversarial loss value is calculated according to the following formula:

[0017] Among them, i NAC is a sample PET image without attenuation correction, For i NAC The mathematical expectation calculated for the sample, D NAC (i NAC ) is the second discriminator pair i NAC The judgment result of i AC is the attenuation-corrected PET image sample, For i AC The mathematical expectation calculated for the sample, D AC (i AC ) is the first discriminator pair i NAC The judgment result of .

[0018] In the PET image attenuation correction model training method of the embodiment of this specification, the cycle consistency loss value is calculated according to the following formula:

[0019] Among them, i NAC is a sample PET image without attenuation correction, For i NAC is the mathematical expectation calculated for the sample, G(i NAC ) is the first generator pair i NAC The attenuation correction result, F(G(i NAC )-i NAC ) is the second generator pair G(i NAC ) with i NAC The reverse transformation result of the difference between ‖F(G(i NAC )-i NAC )‖1 is F(G(i NAC )-i NAC )’s first norm, i AC is the attenuation-corrected PET image sample, For i AC The mathematical expectation of the sample, F(i AC ) is the second generator pair i AC The reverse transformation result, G(F(iAC )-i AC ) is the first generator pair F(i AC ) and i AC The attenuation correction result of the difference, ‖G(F(i AC )-i AC )‖1 is G(F(i AC )-i AC ) is the first norm of .

[0020] In the PET image attenuation correction model training method of the embodiment of this specification, the fidelity loss value is calculated according to the following formula:

[0021] Among them, i AC is the attenuation-corrected PET image sample, For i AC is the mathematical expectation calculated for the sample, G(i AC ) is the first generator pair i AC The attenuation correction result, ‖G(i AC )-i AC ‖1 is G(i AC ) and i AC The first norm of the difference, i NAC is a sample PET image without attenuation correction, For i NAC The mathematical expectation of the sample, F(i NAC ) is the second generator pair i NAC The reverse transformation result, [‖F(i NAC )-i NAC ‖1] is F(i NAC ) and i NAC The first norm of the difference.

[0022] In the PET image attenuation correction model training method of the embodiment of this specification, the part structure loss value is calculated according to the following formula:

[0023] Among them, M is the number of part structure types, N is the number of training samples of the first discriminator, and x jc is the cth target category of sample j of the first discriminator, p jc is the predicted probability of sample j of the first discriminator as the cth target category, R is the number of training samples of the second discriminator, y kc is the cth target category of sample k of the second discriminator, p kc is the predicted probability of sample k of the second discriminator being the c-th target category.

[0024] In the PET image attenuation correction model training method of the embodiment of this specification, after determining the first generator in the first generative adversarial network as a PET image attenuation correction model without CT, the method further includes:

[0025] PET image samples corresponding to structures injected with different PET tracer doses are used to verify the PET image attenuation correction model;

[0026] When the PET image attenuation correction model fails the verification, continuing to train the PET image attenuation correction model;

[0027] When the PET image attenuation correction model passes the verification, the PET image attenuation correction model is retained.

[0028] On the other hand, the embodiments of this specification also provide a method for PET image attenuation correction without CT, including:

[0029] receiving a PET image to be attenuation corrected;

[0030] The PET image to be attenuated corrected is input into a PET image attenuation correction model to obtain an attenuation-corrected PET image.

[0031] On the other hand, the embodiments of this specification also provide a CT-free PET image attenuation correction model training device, comprising:

[0032] A creation module, configured to create a cyclic generative adversarial network; the cyclic generative adversarial network comprising a first generative adversarial network and a second generative adversarial network;

[0033] a training module, configured to train the recurrent generative adversarial network so that the first generative adversarial network learns a mapping relationship of converting an unattenuated PET image into an attenuated PET image, and the second generative adversarial network learns a mapping relationship of converting an attenuation-corrected PET image into an unattenuated PET image;

[0034] A judgment module is used to judge whether the currently trained cyclic generative adversarial network meets the preset conditions;

[0035] A determination module is used to determine the first generator in the first generative adversarial network as a PET image attenuation correction model without CT when the cyclic generative adversarial network obtained by the current training meets a preset condition.

[0036] On the other hand, the embodiments of this specification also provide a PET image attenuation correction device without CT, comprising:

[0037] An image receiving module, configured to receive a PET image to be attenuated corrected;

[0038] The attenuation correction module is used to input the PET image to be attenuated corrected into a PET image attenuation correction model to obtain the attenuation-corrected PET image.

[0039] On the other hand, an embodiment of this specification further provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the computer program executes instructions of the above method when executed by the processor.

[0040] On the other hand, an embodiment of this specification further provides a computer storage medium having a computer program stored thereon, wherein the computer program executes the instructions of the above method when executed by a processor of a computer device.

[0041] On the other hand, an embodiment of this specification further provides a computer program product, which includes a computer program. When the computer program is executed by a processor of a computer device, the computer program executes the instructions of the above method.

[0042] It can be seen from the technical solutions provided in the above embodiments of this specification that the embodiments of this specification use a cyclic generative adversarial network to find the mapping relationship between the un-attenuation-corrected PET image and the attenuation-corrected PET image; since the generator in the trained cyclic generative adversarial network has learned the mapping relationship between the un-attenuation-corrected PET image and the attenuation-corrected PET image; thus, the generator can directly perform attenuation correction processing on the PET image to be attenuated based on the mapping relationship to obtain the attenuation-corrected PET image without referring to the CT image; thereby achieving the goal of reducing or avoiding harmful radiation dose while ensuring the quality of attenuation correction of the PET image. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some of the embodiments described in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without inventive work. In the drawings:

[0044] FIG1 shows a flowchart of a method for training a PET image attenuation correction model without CT in some embodiments of this specification;

[0045] FIG2 shows a schematic diagram of the network structure of a PET image attenuation correction model without CT in some embodiments of this specification;

[0046] FIG3 shows a schematic diagram of the structure of a generator in some embodiments of this specification;

[0047] FIG4 is a schematic diagram showing the structure of a discriminator in some embodiments of this specification;

[0048] FIG5 shows a flowchart of a method for training a PET image attenuation correction model without CT in some other embodiments of this specification;

[0049] FIG6 shows a flowchart of a method for attenuation correction of PET images without CT in some embodiments of this specification;

[0050] FIG7 shows a non-attenuation-corrected PET image of a local tissue structure of a human body in an exemplary embodiment of this specification;

[0051] FIG8 shows a reference PET image of a local tissue structure of a human body in an exemplary embodiment of this specification;

[0052] FIG9 shows an attenuation-corrected PET image of a local tissue structure of a human body in an exemplary embodiment of the present specification;

[0053] FIG10 shows a structural block diagram of a PET image attenuation correction model training device without CT in some embodiments of this specification;

[0054] FIG11 shows a structural block diagram of a PET image attenuation correction device without CT in some embodiments of this specification;

[0055] FIG12 shows a structural block diagram of a computer device in some embodiments of this specification.

[0056] [Explanation of the accompanying drawings] 1011. creation module; 1012. training module; 1013. judgment module; 1014. determination module; 1111. image receiving module; 1112. attenuation correction module; 1202. computer device; 1204. processor; 1206. memory; 1208. driving mechanism; 1210. input / output interface; 1212. input device; 1214. output device; 1216. presentation device; 1218. graphical user interface; 1220. network interface; 1222. communication link; 1224. communication bus. DETAILED DESCRIPTION

[0057] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments derived by those skilled in the art based on the embodiments in this specification without creative effort shall fall within the scope of protection of this specification.

[0058] The embodiments of this specification provide a technical solution for CT-free PET image attenuation correction, which reduces harmful radiation dose while ensuring the quality of PET image attenuation correction. The solution can be applied to attenuation correction processing of PET images of structures in the human or animal body. The structures can be systemic or localized.

[0059] Referring to FIG1 , in some embodiments of this specification, a method for training a PET image attenuation correction model without CT is provided, which may include the following steps:

[0060] Step 101: Create a cycle generative adversarial network (Cycle-GAN); the cycle generative adversarial network includes a first generative adversarial network and a second generative adversarial network.

[0061] For example, in the cyclic generative adversarial network shown in Figure 2, the first generator G and the first discriminator D AC The first generative adversarial network, the second generator F and the second discriminator D can be formed NAC A second generative adversarial network can be formed. The first generator G can realize the conversion of the PET image domain without attenuation correction (Non Attenuation Correction, NAC) to the PET image domain with attenuation correction (Attenuation Correction, AC); the first discriminator D AC The authenticity of the attenuation-corrected PET image output by the first generator G can be judged, and the structure of the part to which the image belongs can be identified. The second generator F can realize the reverse conversion from the attenuation-corrected PET image domain to the non-attenuation-corrected PET image domain. The first discriminator D AC The authenticity of the attenuation-corrected PET image output by the first generator G can be determined and the structure of the part to which the image belongs can be confirmed.

[0062] In some embodiments of the present specification, the first generator G and the second generator F may have the same encoder-decoder structure. Referring to FIG3 , the encoder of each generator may include four basic units, each of which includes a convolution layer with a convolution kernel size of 3×3 and a stride of 2, a normalization layer, and a ReLU activation function. The decoder structure of each generator also includes four basic units, each of which includes a transposed deconvolution layer with a convolution kernel size of 3×3 and a stride of 2, a normalization layer, and a ReLU activation function.

[0063] As shown in Figure 3, in the generator, when the network (i.e., the generator's neural network) receives an input image (i.e., a PET image), it first passes through a convolutional layer, increasing the image dimension from 1 to 32. It then passes through a 3x3 convolutional unit block with a stride of 2. This convolution downsamples the image, reducing its dimensionality while extracting features from the data. A normalization layer is used to normalize the input data in deep neural networks. This layer calculates the mean and standard deviation of each batch of data and normalizes the data, making the network more robust to variations in the input data. This normalization layer not only accelerates model training but also provides regularization, helping to prevent overfitting. The ReLU function has the characteristic of sparse activation, meaning that only some neurons are activated (outputting non-zero values) while others remain inactive. This helps the model more effectively learn and represent key features of the input data, reduces model complexity, and improves computational efficiency. After passing through four such units, the input image is gradually restored to its original size using transposed deconvolution. The role of the final convolutional layer is to restore the feature dimension from 32 to 1.

[0064] In some embodiments of this specification, the first discriminator D AC and the second discriminator D NAC The same structure can also be used. As shown in Figure 4, the encoder of each discriminator can include four basic units. Each basic unit can include a convolutional layer with a kernel size of 4×4 and a stride of 1, a normalization layer, and a Leaky ReLU activation function. Three fully connected layers are attached after the four basic units to output the true or false discrimination and part structure discrimination of the image.

[0065] When the discriminator receives network input, it first uses convolution to increase the feature dimension from 1 to 32. Then, the convolution within the four basic units also downsamples the image. In deep learning, flattening refers to converting multidimensional input data into a one-dimensional form, usually performed before the input layer of the neural network. This process can be understood as expanding the structure of the input data into a long vector, so that the data can be input into the fully connected layer or other layers requiring one-dimensional input. Then, two fully connected layers are constructed: one to output a judgment on the authenticity of the image, and the other to output a judgment on the structure of the image part.

[0066] In the initially created cyclic generative adversarial network, since the parameters of the first and second generative adversarial networks are randomly assigned, they cannot meet the application requirements of PET attenuation correction without CT. Therefore, the first and second generative adversarial networks need to be trained to obtain a cyclic generative adversarial network that meets the preset conditions.

[0067] Step 102: Train the recurrent generative adversarial network so that the first generative adversarial network learns a mapping relationship for converting an unattenuated PET image into an attenuated PET image, and the second generative adversarial network learns a mapping relationship for converting an attenuated PET image into an unattenuated PET image.

[0068] In some embodiments of the present specification, image pairs and corresponding body part structure labels can be extracted from datasets of unattenuation-corrected and attenuation-corrected PET images to construct datasets (including training, test, and validation sets) required for model training. Furthermore, when training the recurrent generative adversarial network, an Adam optimizer can be used for model optimization.

[0069] Step 103: Determine whether the cyclic generative adversarial network obtained by the current training meets the preset conditions; if the cyclic generative adversarial network obtained by the current training meets the preset conditions, execute step 104; otherwise, execute step 102.

[0070] In some embodiments of this specification, determining whether the currently trained cyclic generative adversarial network meets the preset conditions may include: according to the formula L=L adv1 +L adv2 +λ1L cyc +λ2L ide +λ3L site1 +λ4L site2 Calculate the total loss value of the currently trained cyclic generative adversarial network; then determine whether the total loss value is less than a loss threshold; when the total loss value is less than the loss threshold, confirm that the currently trained cyclic generative adversarial network meets the preset conditions; otherwise, confirm that the currently trained cyclic generative adversarial network does not meet the preset conditions.

[0071] Among them, L is the total loss value, L adv1 is the adversarial loss value of the first generator in the first generative adversarial network, L adv2 is the adversarial loss value of the second generator in the second generative adversarial network, L cyc is the cycle consistency loss value of the first generator and the second generator, L ide is the fidelity loss value of the first generator and the second generator, L site1 is the part structure loss of the first discriminator in the first generative adversarial network, L site2 is the part structure loss of the second discriminator in the second generative adversarial network, λ1, λ2, λ3 and λ4 correspond to L cyc 、L ide 、L site1 and L site2 The coefficient of .

[0072] Adversarial loss is used to encourage the generator to generate realistic images in the target domain to deceive the discriminator network. At the same time, the discriminator strives to distinguish between generated images and real images. This adversarial training promotes competition between the generator and the discriminator, which in turn helps the generator produce more convincing conversion results. Therefore, adversarial loss is a loss that should be considered in generative adversarial networks.

[0073] In some embodiments of this specification, the adversarial loss value is calculated according to the following formula:

[0074] Among them, i NAC is a sample PET image without attenuation correction, For i NAC The mathematical expectation calculated for the sample, D NAC (i NAC ) is the second discriminator pair i NAC The judgment result of i AC is the attenuation-corrected PET image sample, For i AC The mathematical expectation calculated for the sample, D AC (i AC ) is the first discriminator pair i NAC The judgment result of .

[0075] Using adversarial loss alone cannot ensure that the mapping function successfully transforms the source input into the target output. Therefore, in order to promote i NAC and F(G(i NAC )) consistency, in the embodiment of this specification, a cycle consistency loss is introduced; wherein G(i NAC ) represents the first generator pair i NAC The attenuation correction result (ie, the attenuation-corrected PET image), F(G(i NAC )) represents the NAC ) is the reverse conversion of .

[0076] In some embodiments of this specification, the cycle consistency loss value can be calculated according to the formula Calculated. Among them, i NAC is a sample PET image without attenuation correction, For i NAC is the mathematical expectation calculated for the sample, G(i NAC ) is the first generator pair i NAC The attenuation correction result, F(G(i NAC )-i NAC ) is the second generator pair G(i NAC) and i NAC The reverse transformation result of the difference between ‖F(G(i NAC )-i NAC )‖1 is F(G(i NAC )-i NAC )’s first norm, i AC is the attenuation-corrected PET image sample, For i AC The mathematical expectation of the sample, F(i AC ) is the second generator pair i AC The reverse transformation result, G(F(i AC )-i AC ) is the first generator pair F(i AC ) and i AC The attenuation correction result of the difference, ‖G(F(i AC )-i AC )‖1 is G(F(i AC )-i AC ) is the first norm of .

[0077] In order to preserve the characteristics of the source image, Cycle-GAN introduces a fidelity loss L ide , which encourages the generator to not only perform effective image transformation but also maintain the characteristics of the image, hoping that G(i AC ) and i AC As close as possible.

[0078] In some embodiments of this specification, the fidelity loss value can be calculated according to the formula Calculated; where i AC is the attenuation-corrected PET image sample, For i AC is the mathematical expectation calculated for the sample, G(i AC ) is the first generator pair i AC The attenuation correction result, ‖G(i AC )-i AC ‖1 is G(i AC ) and i AC The first norm of the difference, i NAC is a sample PET image without attenuation correction, For i NAC The mathematical expectation of the sample, F(i NAC ) is the second generator pair i NAC The reverse transformation result, [‖F(i NAC )-i NAC ‖1] is F(i NAC ) and i NAC The first norm of the difference.

[0079] For structures within the same location, they exhibit significant similarity in shape and size. However, for different locations, their appearance varies significantly, demonstrating distinct specificity. Therefore, to maximize the specificity of different tissue structures and enhance image contrast, location differences can be incorporated into network design as a priori information, i.e., incorporating location structural differences into the loss function.

[0080] In some embodiments of this specification, the site structure loss value can be calculated according to the following formula:

[0081] Among them, M is the number of part structure types, N is the number of training samples of the first discriminator, and x jc is the cth target category of sample j of the first discriminator, p jc is the predicted probability of sample j of the first discriminator as the cth target category, R is the number of training samples of the second discriminator, y kc is the cth target category of sample k of the second discriminator, p kc is the predicted probability of sample k of the second discriminator being the c-th target category.

[0082] Step 104: When the recurrent generative adversarial network obtained through the current training meets a preset condition, the first generator in the first generative adversarial network is determined as a PET image attenuation correction model without CT.

[0083] In the embodiments of the present specification, a mapping relationship between an uncorrected PET image and an attenuation-corrected PET image is found through a cyclic generative adversarial network. Since the generator in the trained cyclic generative adversarial network has learned the mapping relationship between the uncorrected PET image and the attenuation-corrected PET image, the generator can directly perform attenuation correction processing on the PET image to be corrected for attenuation based on the mapping relationship to obtain the attenuation-corrected PET image without referring to the CT image. This achieves the goal of reducing or avoiding harmful radiation dose while ensuring the quality of attenuation correction of the PET image.

[0084] Referring to FIG5 , some embodiments of this specification also provide another PET image attenuation correction model training method, which includes the following steps:

[0085] Step 501: Create a cycle generative adversarial network (Cycle-GAN); the cycle generative adversarial network includes a first generative adversarial network and a second generative adversarial network.

[0086] Step 502: Train the recurrent generative adversarial network so that the first generative adversarial network learns a mapping relationship for converting an unattenuated PET image into an attenuated PET image, and the second generative adversarial network learns a mapping relationship for converting an attenuated PET image into an unattenuated PET image.

[0087] Step 503: Determine whether the cyclic generative adversarial network obtained by the current training meets the preset conditions; if the cyclic generative adversarial network obtained by the current training meets the preset conditions, execute step 504; otherwise, execute step 502.

[0088] Step 504: When the recurrent generative adversarial network obtained through current training meets a preset condition, the first generator in the first generative adversarial network is determined as a PET image attenuation correction model without CT.

[0089] Step 505: Validate the PET image attenuation correction model using PET image samples corresponding to structures at sites injected with different PET tracer doses.

[0090] Step 506: Determine whether the PET image attenuation correction model passes the verification. If so, proceed to step 507; otherwise, proceed to step 502 (i.e., continue training the PET image attenuation correction model). Passing the verification here may mean that the total pass rate of the verification samples with different tracer doses reaches a set threshold; correspondingly, failing the verification may mean that the total pass rate of the verification samples with different tracer doses does not reach the set threshold.

[0091] Step 507: retain the PET image attenuation correction model.

[0092] The retained PET image attenuation correction model can be applied to actual PET image attenuation correction processing scenarios.

[0093] Compared with the model training method shown in FIG1 , the model training method shown in FIG5 adds a step of verifying the PET image attenuation correction model using PET image samples corresponding to the local structures injected with different PET tracer doses. That is, verifying whether the trained model can attenuate correct the PET images obtained at different tracer dose levels into PET images with complete details preserved. If the verification is passed, it indicates that the trained PET image attenuation correction model is feasible and can be actually applied to the attenuation correction processing of PET images.

[0094] Referring to FIG6 , based on the PET image attenuation correction model obtained by the above-mentioned model training method, an embodiment of this specification provides a CT-free PET image attenuation correction method, which may include the following steps:

[0095] Step 601: receiving a PET image to be attenuated corrected;

[0096] Step 602: Input the PET image to be attenuation corrected into a PET image attenuation correction model to obtain an attenuation-corrected PET image.

[0097] For example, in an exemplary embodiment of the present specification, assuming that an image of a certain human body structure without attenuation correction is shown in Figure 7, based on the traditional PET image attenuation correction with the help of CT images, the PET image shown in Figure 7 can be attenuated and corrected to a PET image as shown in Figure 8; and based on the PET image attenuation correction method of the embodiment of the present specification, the PET image shown in Figure 7 can be attenuated and corrected to a PET image as shown in Figure 9; by comparing Figures 8 and 9, it can be seen that both Figures 8 and 9 can obtain PET images that are clearer and more structurally complete than Figure 7, but the PET image attenuation correction method based on the embodiment of the present specification effectively avoids or reduces the harmful radiation dose to the human body because it does not refer to the CT image.

[0098] Although the process flows described above include multiple operations occurring in a particular order, it should be understood that these processes may include more or fewer operations, which may be performed sequentially or in parallel (eg, using parallel processors or a multi-threaded environment).

[0099] Corresponding to the above-mentioned method for training a PET image attenuation correction model without CT, embodiments of this specification further provide a device for training a PET image attenuation correction model without CT. Referring to FIG10 , in some embodiments of this specification, the device for training a PET image attenuation correction model without CT may include:

[0100] A creation module 1011 is configured to create a cyclic generative adversarial network, wherein the cyclic generative adversarial network includes a first generative adversarial network and a second generative adversarial network.

[0101] A training module 1012 is configured to train the recurrent generative adversarial network so that the first generative adversarial network learns a mapping relationship for converting an unattenuated PET image into an attenuated PET image, and the second generative adversarial network learns a mapping relationship for converting an attenuation-corrected PET image into an unattenuated PET image;

[0102] A judgment module 1013 is used to judge whether the cyclic generative adversarial network obtained by the current training meets the preset conditions;

[0103] The determination module 1014 is configured to determine the first generator in the first generative adversarial network as a PET image attenuation correction model without CT when the currently trained cyclic generative adversarial network meets a preset condition.

[0104] Corresponding to the above-mentioned PET image attenuation correction method without CT, embodiments of this specification further provide a PET image attenuation correction device without CT. Referring to FIG11 , in some embodiments of this specification, the PET image attenuation correction device without CT may include:

[0105] An image receiving module 1111 is configured to receive a PET image to be attenuated corrected;

[0106] The attenuation correction module 1112 is configured to input the PET image to be attenuated corrected into a PET image attenuation correction model to obtain an attenuation-corrected PET image.

[0107] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0108] It should be noted that in the embodiments of this specification, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved are all information and data authorized by the user and fully authorized by all parties.

[0109] The embodiments of this specification also provide a computer device. As shown in Figure 12, in some embodiments of this specification, the computer device 1202 may include one or more processors 1204, such as one or more central processing units (CPUs) or graphics processing units (GPUs), and each processing unit may implement one or more hardware threads. The computer device 1202 may also include any memory 1206, which is used to store any type of information such as code, settings, data, etc. In a specific embodiment, the computer program on the memory 1206 and capable of running on the processor 1204, when the computer program is run by the processor 1204, can execute the instructions of the method described in any of the above embodiments. Non-limitingly, for example, the memory 1206 may include any one or more combinations of the following: any type of RAM, any type of ROM, a flash memory device, a hard disk, an optical disk, etc. More generally, any memory can use any technology to store information. Further, any memory can provide volatile or non-volatile retention of information. Further, any memory can represent a fixed or removable component of the computer device 1202. In one embodiment, when the processor 1204 executes the associated instructions stored in any memory or combination of memories, the computer device 1202 can perform any operation of the associated instructions. The computer device 1202 also includes one or more drive mechanisms 1208 for interacting with any memory, such as a hard disk drive mechanism, an optical disk drive mechanism, etc.

[0110] The computer device 1202 may also include an input / output interface 1210 (I / O) for receiving various inputs (via input devices 1212) and for providing various outputs (via output devices 1214). A specific output mechanism may include a presentation device 1216 and an associated graphical user interface 1218 (GUI). In other embodiments, the input / output interface 1210 (I / O), input devices 1212, and output devices 1214 may not be included, and the computer device 1202 may simply be a computer device in a network. The computer device 1202 may also include one or more network interfaces 1220 for exchanging data with other devices via one or more communication links 1222. One or more communication buses 1224 couple the components described above together.

[0111] The communication link 1222 may be implemented in any manner, for example, via a local area network, a wide area network (e.g., the Internet), a point-to-point connection, etc., or any combination thereof. The communication link 1222 may include any combination of hardwired links, wireless links, routers, gateway functions, name servers, etc., governed by any protocol or combination of protocols.

[0112] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), computer-readable storage media, and computer program products of some embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processor to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processor produce a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0113] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processor to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0114] These computer program instructions can also be loaded onto a computer or other programmable data processor so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0115] In a typical configuration, a computer device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0116] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0117] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computer device. As defined in this specification, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0118] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the embodiments of this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0119] Embodiments of this specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. Embodiments of this specification may also be practiced in distributed computing environments where tasks are performed by remote processors connected via a communications network. In distributed computing environments, program modules may be located in local and remote computer storage media, including storage devices.

[0120] It should also be understood that in the embodiments of this specification, the term "and / or" is merely a description of the association relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0121] The various embodiments in this specification are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are generally similar to the method embodiments, so the description is relatively simple. For relevant parts, refer to the description of the method embodiments.

[0122] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiments of this specification. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0123] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A method for training a PET image attenuation correction model without CT, characterized in that: include: Creating recurrent generative adversarial networks; The cyclic generative adversarial network includes a first generative adversarial network and a second generative adversarial network; Training the recurrent generative adversarial network so that the first generative adversarial network learns a mapping relationship of converting an unattenuated corrected PET image into an attenuated corrected PET image, and the second generative adversarial network learns a mapping relationship of converting an attenuated corrected PET image into an unattenuated corrected PET image; Determine whether the cyclic generative adversarial network obtained by the current training meets the preset conditions; When the recurrent generative adversarial network obtained through the current training meets a preset condition, the first generator in the first generative adversarial network is determined as a PET image attenuation correction model without CT.

2. The PET image attenuation correction model training method according to claim 1, characterized in that: Determine whether the currently trained cyclic generative adversarial network meets the preset conditions, including: According to the formula L = L adv1 +L adv2 +λ1L cyc +λ2L ide +λ3L site1 +λ4L site2 Calculate the total loss value of the current training cyclic generative adversarial network; Determining whether the total loss value is less than a loss threshold; When the total loss value is less than the loss threshold, it is confirmed that the cyclic generative adversarial network obtained by the current training meets the preset condition; otherwise, it is confirmed that the cyclic generative adversarial network obtained by the current training does not meet the preset condition; Among them, L is the total loss value, L adv1 is the adversarial loss value of the first generator in the first generative adversarial network, L adv2 is the adversarial loss value of the second generator in the second generative adversarial network, L cyc is the cycle consistency loss value of the first generator and the second generator, L ide is the fidelity loss value of the first generator and the second generator, L site1 is the part structure loss of the first discriminator in the first generative adversarial network, L site2 is the part structure loss of the second discriminator in the second generative adversarial network, and λ1, λ2, λ3 and λ4 correspond to L cyc , L ide , L site1 and L site2 The coefficient of .

3. The PET image attenuation correction model training method according to claim 2, characterized in that: The adversarial loss value is calculated according to the following formula: Among them, i NAC is a sample PET image without attenuation correction, For i NAC The mathematical expectation calculated for the sample, D NAC (i NAC ) is the second discriminator pair i NAC The judgment result of i AC is a sample of attenuation-corrected PET images, For i AC The mathematical expectation calculated for the sample, D AC (i AC ) is the first discriminator pair i NAC The discrimination result of .

4. The PET image attenuation correction model training method according to claim 2, characterized in that: The cycle consistency loss value is calculated according to the following formula: Among them, i NAC is a sample PET image without attenuation correction, For i NAC is the mathematical expectation calculated for the sample, G(i NAC ) is the first generator pair i NAC The attenuation correction result, F(G(i NAC )-i NAC ) is the second generator pair G(i NAC ) and i NAC The reverse transformation result of the difference between NAC )-i NAC )‖1 is F(G(i NAC )-i NAC ), i AC is a sample of attenuation-corrected PET images, For i AC is the mathematical expectation calculated for the sample, F(i AC ) is the second generator pair i AC The reverse transformation result, G(F(i AC )-i AC ) is the first generator pair F(i AC ) and i AC The attenuation correction result of the difference, ‖G(F(i AC )-i AC )‖1 is G(F(i AC )-i AC ) is the first norm of .

5. The PET image attenuation correction model training method according to claim 2, characterized in that: The fidelity loss value is calculated according to the following formula: Among them, i AC is a sample of attenuation-corrected PET images, For i AC is the mathematical expectation calculated for the sample, G(i AC ) is the first generator pair i AC The attenuation correction result, ‖G(i AC )-i AC ‖1 is G(i AC ) and i AC The first norm of the difference, i NAC is a sample PET image without attenuation correction, For i NAC is the mathematical expectation calculated for the sample, F(i NAC ) is the second generator pair i NAC The reverse transformation result, [‖F(i NAC )-i NAC ‖1] is F(i NAC ) and i NAC The first norm of the difference.

6. The PET image attenuation correction model training method according to claim 2, characterized in that: The structural loss value of the part is calculated according to the following formula: Where M is the number of part structure types, N is the number of training samples for the first discriminator, and x jc is the cth target category of sample j of the first discriminator, p jc is the predicted probability of sample j of the first discriminator as the cth target category, R is the number of training samples of the second discriminator, y kc is the cth target category of sample k of the second discriminator, p kc is the predicted probability that sample k of the second discriminator is the c-th target category.

7. The PET image attenuation correction model training method according to claim 1, characterized in that: After determining the first generator in the first generative adversarial network as a PET image attenuation correction model without CT, the method further includes: Using PET image samples corresponding to the structures of the sites injected with different PET tracer doses, respectively, to verify the PET image attenuation correction model; When the PET image attenuation correction model fails the verification, the PET image attenuation correction is continued. The model is trained; When the PET image attenuation correction model passes the verification, the PET image attenuation correction model is retained.

8. A PET image attenuation correction method without CT, characterized in that: include: receiving a PET image to be attenuation corrected; The PET image to be attenuation corrected is input into a PET image attenuation correction model to obtain an attenuation-corrected PET image.

9. A PET image attenuation correction model training device without CT, characterized in that: include: Create a module for creating recurrent generative adversarial networks; The cyclic generative adversarial network includes a first generative adversarial network and a second generative adversarial network; A training module, used for training the recurrent generative adversarial network, so that the first generative adversarial network learns a mapping relationship of converting an unattenuated PET image into an attenuated PET image, and the second generative adversarial network learns a mapping relationship of converting an attenuated PET image into an unattenuated PET image; A judgment module is used to judge whether the recurrent generative adversarial network obtained by the current training meets the preset conditions; A determination module is used to determine the first generator in the first generative adversarial network as a PET image attenuation correction model without CT when the recurrent generative adversarial network obtained by current training meets a preset condition.

10. A PET image attenuation correction device without CT, characterized in that: include: An image receiving module, used for receiving a PET image to be attenuated corrected; The attenuation correction module is used to input the PET image to be attenuation corrected into the PET image attenuation correction model to obtain the PET image after attenuation correction.

11. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: When the computer program is executed by the processor, the computer program executes the instructions of the method according to any one of claims 1 to 8.

12. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor of a computer device, the computer program executes the instructions of the method according to any one of claims 1 to 8.

13. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor of a computer device, the computer program executes instructions of the method according to any one of claims 1 to 8.

Images