PET image self-attenuation correction method and system, medium and product

Abstract

The invention provides a PET image self-attenuation correction method and system, a medium and a product, and the method employs a generation strategy from coarse to fine, is based on a PET self-attenuation correction technology of a dual-branch residual diffusion model, enables a PET image and a CT image to be mutually guided and cooperatively optimized in a generation process through a dual-branch mutual guiding module and a focus perception refining module, and improves the precision of the PET image and the CT image. In combination with physical prior constraints, the dependence on synthetic CT is reduced, error propagation and amplification are effectively inhibited, and the precision, efficiency and clinical applicability of attenuation correction are improved.

Description

Technical Field

[0001] This application relates to the fields of medical image processing and artificial intelligence technology, and in particular to a method, system, medium and product for PET image self-attenuation correction. Background Technology

[0002] Positron emission tomography (PET) is a functional molecular imaging technique that obtains information on the metabolic activities of human tissues by detecting the distribution of radioactive tracers injected into the body. It plays an irreplaceable role in fields such as tumor diagnosis, assessment of neurological diseases, and cardiac function examination, and provides important evidence for disease diagnosis and treatment decisions.

[0003] However, the physical process of PET imaging has inherent limitations: gamma photons (energy 511 keV) generated by the decay of radioactive tracers are attenuated by absorption and scattering in human tissue as they penetrate, resulting in a weakened signal received by the detector. If this attenuation effect is not corrected, it will severely affect the quality and quantitative accuracy of PET images, especially the calculation of the standard uptake value (SUV). As a core indicator of PET quantitative analysis, the SUV is a crucial basis for differentiating between benign and malignant tumors, evaluating treatment efficacy, and monitoring recurrence; its calculation error can directly mislead clinical diagnostic results.

[0004] Currently, the gold standard for PET attenuation correction in the clinical field is to combine it with computed tomography (CT) scans. The core principle of this approach is to leverage the precise anatomical structure depicted by CT scans to convert the acquired CT images into a 511keV gamma photon attenuation coefficient map relevant to PET imaging, and then perform attenuation correction on the raw PET data based on this attenuation coefficient map. However, this method exposes patients to additional ionizing radiation and increases scan time and medical costs.

[0005] With the research and development of artificial intelligence in the field of medical imaging, deep learning methods have been introduced to achieve PET self-attenuation correction, that is, generating attenuation-corrected PET images (AC-PET images) only from PET images before attenuation correction (NAC-PET images).

[0006] Existing deep learning-based methods for PET self-attenuation correction mainly fall into two categories:

[0007] (1) Direct attenuation correction method: This type of method uses an end-to-end deep learning model, directly using NAC-PET images as input and AC-PET images as output for model training, to achieve direct conversion from NAC-PET to AC-PET. This method has a simple process, but due to the lack of clear human anatomy information guidance—because NAC-PET images themselves only reflect the metabolic activity of tissues and cannot provide fine anatomical details, the model has difficulty accurately recovering the attenuation characteristics of different tissues, and the final AC-PET images generated are overly smooth with limited quantitative accuracy, which is difficult to meet clinical requirements.

[0008] (2) Indirect attenuation correction method. This type of method adopts a two-step serial technical process. In the first step, a pseudo-CT image reflecting human anatomy is synthesized from the NAC-PET image using a deep learning model. In the second step, the synthesized pseudo-CT image is converted into an attenuation coefficient map, and then the NAC-PET image is attenuated based on the attenuation coefficient map to obtain the AC-PET image. This type of method supplements anatomical information by introducing pseudo-CT images, but its performance is limited by two reasons:

[0009] ① Cross-modal synthesis is difficult: NAC-PET images and CT images are image data of different modalities. The former focuses on functional metabolic information, while the latter focuses on anatomical density information. The modal differences between the two are huge, and NAC-PET images themselves lack fine anatomical details, making it extremely difficult to synthesize high-quality, high-fidelity pseudo-CT images directly from NAC-PET images.

[0010] ② Error accumulation and optimization isolation: The two steps of this type of method (pseudo-CT synthesis and pseudo-CT-based attenuation correction) are independent serial processes. The model training process is isolated from each other, and information exchange and collaborative optimization cannot be achieved. Any small error generated in the pseudo-CT synthesis stage will be transmitted and amplified in the subsequent attenuation correction step, which will ultimately seriously affect the quality and quantitative accuracy of AC-PET images.

[0011] Furthermore, the model architectures used in existing deep learning self-attenuation correction methods are mostly traditional convolutional neural networks (CNNs) or generative adversarial networks (GANs). Traditional convolutional neural networks (CNNs) have limited ability to capture image details, which can easily lead to over-smoothing of the generated AC-PET images. Generative adversarial networks (GANs) are prone to pattern collapse, which means that the generated images lack diversity and cannot accurately match the individual anatomical differences and metabolic characteristics of different patients, making it difficult to meet the clinical requirements for high-precision quantification.

[0012] Therefore, developing a new PET self-attenuation correction method that can accurately synthesize anatomical structures, achieve joint optimization of PET and CT, and ensure high fidelity of the generated images has significant clinical value and technical challenges. Summary of the Invention

[0013] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a PET image self-attenuation correction method, system, medium and product to solve the technical problems of insufficient correction accuracy, inaccurate quantification and error accumulation in the prior art.

[0014] To achieve the above and other related objectives, a first aspect of this application provides a PET image self-attenuation correction method, comprising: constructing a PET self-attenuation correction model; the PET self-attenuation correction model including a two-branch residual diffusion model; generating an attenuation-corrected PET target image based on an acquired unattenuated real-time PET image and the PET self-attenuation correction model; wherein the construction method of the PET self-attenuation correction model includes: constructing a training dataset; the training dataset including multiple unattenuated PET historical images, attenuated PET historical images corresponding to each unattenuated PET historical image, and CT historical images; inputting each unattenuated PET historical image, the attenuated PET historical image corresponding to each unattenuated PET historical image, and the CT historical images into the two-branch residual diffusion model, performing a two-branch learning operation, and constructing a primary PET self-attenuation correction model based on a preset total loss function; performing iterative training on the primary PET self-attenuation correction model to update the total loss function, and adjusting and optimizing the parameters of the primary PET self-attenuation correction model according to the updated total loss function to generate a finally converged PET self-attenuation correction model.

[0015] In some embodiments of the first aspect of this application, the dual-branch residual diffusion model includes a PET branch residual diffusion model and a CT branch residual diffusion model; wherein, the method of constructing the primary PET self-attenuation correction model includes: inputting the unattenuated corrected PET historical image and the corresponding attenuated corrected PET historical image into the PET branch residual diffusion model, performing a PET branch learning operation to generate a coarse attenuated corrected PET historical image corresponding to the unattenuated corrected PET historical image; inputting the unattenuated corrected PET historical image and the corresponding CT historical image into the PET branch residual diffusion model. In the CT branch residual diffusion model, a CT branch learning operation is performed to generate a coarse reconstructed CT historical image corresponding to the unattenuated corrected PET historical image. Based on the coarse attenuated corrected PET historical image and the coarse reconstructed CT historical image, a feature fusion operation and a feature decoding operation are performed to generate a fine attenuated corrected PET historical image and a fine CT reconstructed image corresponding to the unattenuated corrected PET historical image. Based on the fine attenuated corrected PET historical image and the fine CT reconstructed image, the total loss function is calculated to construct a primary PET self-attenuation correction model.

[0016] In some embodiments of the first aspect of this application, the PET self-attenuation correction model further includes a dual-branch mutual guidance module; wherein, the method of performing feature fusion and feature decoding operations based on the coarse attenuation-corrected PET historical image and coarse reconstructed CT historical image corresponding to the unattenuated PET historical image to generate a fine attenuation-corrected PET historical image and a fine CT reconstructed image corresponding to the unattenuated PET historical image includes: performing a feature fusion operation based on the dual-branch mutual guidance module and based on the coarse attenuation-corrected PET historical image and coarse reconstructed CT historical image corresponding to the unattenuated PET historical image to generate PET branch fusion historical features and CT branch fusion historical features; performing a first decoding operation based on the PET branch residual diffusion model and based on the PET branch fusion historical features to generate a fine attenuation-corrected PET historical image corresponding to the unattenuated PET historical image; and performing a second decoding operation based on the CT branch residual diffusion model and based on the CT branch fusion historical features to generate a fine CT reconstructed image corresponding to the unattenuated PET historical image.

[0017] In some embodiments of the first aspect of this application, the method of generating PET branch fusion historical features and CT branch fusion historical features based on the dual-branch mutual guidance module and according to the coarse attenuation-corrected PET historical image and coarse reconstructed CT historical image corresponding to the unattenuated PET historical image includes: performing a first denoising feature extraction operation on the coarse attenuation-corrected PET historical image corresponding to the unattenuated PET historical image based on the PET branch residual diffusion model to generate PET branch interaction historical features; performing a second denoising feature extraction operation on the coarse reconstructed CT historical image corresponding to the unattenuated PET historical image based on the CT branch residual diffusion model to generate CT branch interaction historical features; performing feature transformation operations on the PET branch interaction historical features and CT branch interaction historical features respectively based on the linear layer of the dual-branch mutual guidance module to generate PET branch attention historical features and CT branch attention historical features; and performing feature fusion operations on the PET branch attention historical features and CT branch attention historical features respectively based on the cross-attention layer of the dual-branch mutual guidance module to generate PET branch fusion historical features and CT branch fusion historical features.

[0018] In some embodiments of the first aspect of this application, the PET branch attention history features include PET branch query history features, PET branch key history features, and PET branch value history features; the CT branch attention history features include CT branch query history features, CT branch key history features, and CT branch value history features; wherein, based on the cross-attention layer of the dual-branch mutual guidance module, the method of performing feature fusion operations on the PET branch attention history features and the CT branch attention history features respectively to generate PET branch fusion history features and CT branch fusion history features includes: based on the cross-attention layer of the dual-branch mutual guidance module, performing a first multi-head attention calculation operation on the PET branch query history features, the CT branch key history features, and the CT branch value history features, and simultaneously performing a second multi-head attention calculation operation on the CT branch query history features, the PET branch key history features, and the PET branch value history features to generate PET branch fusion history features and CT branch fusion history features.

[0019] In some embodiments of the first aspect of this application, the PET self-attenuation correction model further includes a lesion perception refinement module; wherein, the method of generating a fine attenuation-corrected PET historical image corresponding to an unattenuated PET historical image by performing a first decoding operation based on the PET branch residual diffusion model and according to the PET branch fusion history features includes: performing a feature segmentation operation on the coarse attenuation-corrected PET historical image corresponding to the unattenuated PET historical image based on the lesion perception refinement module to generate a lesion history mask; and performing a first decoding operation on the PET branch fusion history features based on the PET branch residual diffusion model and according to the lesion history mask to generate a fine attenuation-corrected PET historical image corresponding to the unattenuated PET historical image.

[0020] In some embodiments of the first aspect of this application, the method of constructing a primary PET self-attenuation correction model by calculating a total loss function based on the finely attenuated corrected PET historical image and the finely reconstructed CT image corresponding to the unattenuated corrected PET historical image includes: calculating a loss function of the PET branch residual diffusion model based on the finely attenuated corrected PET historical image corresponding to the unattenuated corrected PET historical image; calculating a loss function of the CT branch residual diffusion model based on the finely reconstructed CT image corresponding to the unattenuated corrected PET historical image; inputting the unattenuated corrected PET historical image and the finely reconstructed CT image corresponding to the unattenuated corrected PET historical image into a pre-constructed attenuation prior module to generate an attenuated corrected PET reference image; calculating an attenuation prior constraint loss function based on the attenuated corrected PET reference image and the finely attenuated corrected PET historical image corresponding to the unattenuated corrected PET historical image; and calculating a total loss function based on the loss function of the PET branch residual diffusion model, the loss function of the CT branch residual diffusion model, and the attenuation prior constraint loss function, thereby constructing a primary PET self-attenuation correction model.

[0021] To achieve the above and other related objectives, a second aspect of this application provides a PET image self-attenuation correction system, comprising: a model building module for building a PET self-attenuation correction model; the PET self-attenuation correction model including a two-branch residual diffusion model; and a correction module for generating an attenuation-corrected PET target image based on an acquired unattenuated real-time PET image and the PET self-attenuation correction model; wherein the PET self-attenuation correction model is built by constructing a training dataset; the training dataset including multiple unattenuated historical PET images and corresponding images for each unattenuated historical PET image. The attenuated corrected PET historical images and CT historical images are used; each of the unattenuated corrected PET historical images, the corresponding attenuated corrected PET historical images, and the CT historical images are input into a two-branch residual diffusion model, a two-branch learning operation is performed, and a primary PET self-attenuation correction model is constructed based on a preset total loss function; the primary PET self-attenuation correction model is iteratively trained to update the total loss function, and the parameters of the primary PET self-attenuation correction model are adjusted and optimized according to the updated total loss function to generate a finally converged PET self-attenuation correction model.

[0022] To achieve the above and other related objectives, a third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the PET image self-attenuation correction method as described above.

[0023] To achieve the above and other related objectives, a fourth aspect of this application provides a computer program product comprising computer program code that, when executed on a computer, causes the computer to implement the PET image self-attenuation correction method as described above.

[0024] As described above, the PET image self-attenuation correction method, system, medium, and product of this application have the following beneficial effects:

[0025] (1) No additional CT scan required: The method in this application only requires NAC-PET images to complete attenuation correction, which reduces radiation risk and medical costs and avoids radiation exposure and economic burden caused by additional CT scans.

[0026] (2) Improved calibration accuracy and quantitative fidelity: Through the dual-branch mutual guidance module and the lesion perception refinement module, PET images and CT images guide each other and optimize in collaboration during the generation process. Combined with physical prior constraints, the dependence on synthetic CT is reduced, the propagation and amplification of errors are effectively suppressed, and the accuracy and efficiency of attenuation correction are improved, so that the model can generate AC-PET images with clearer anatomical details and more accurate lesion SUV values.

[0027] (3) High-fidelity image generation: The residual diffusion model is used as the backbone network. Compared with CNN or GAN, it can better capture complex data distribution and generate clearer and more detailed images, avoiding problems such as over-smoothing and mode collapse.

[0028] (4) Fully automated end-to-end process, simple and efficient operation: After the model training is completed, end-to-end self-attenuation correction can be achieved. The operation is simple and efficient, without manual intervention, and easy to clinical application and promotion. Attached Figure Description

[0029] Figure 1 The diagram shown is a flowchart of a PET image self-attenuation correction method in one embodiment of this application.

[0030] Figure 2 The diagram shown is a flowchart illustrating the process of constructing a primary PET self-attenuation correction model in one embodiment of this application.

[0031] Figure 3 The diagram shown is a schematic representation of the training process for coarse attenuation correction in one embodiment of this application.

[0032] Figure 4 The diagram shown is a schematic representation of the training process for fine attenuation correction in one embodiment of this application.

[0033] Figure 5 The diagram shown is a structural schematic of a dual-branch mutual guidance module in one embodiment of this application.

[0034] Figure 6 The diagram shown is a structural schematic of a lesion perception refinement module in one embodiment of this application.

[0035] Figure 7 The diagram shown is a structural schematic of the attenuation prior module in one embodiment of this application.

[0036] Figure 8 The diagram shown is a flowchart illustrating the process of constructing a PET self-attenuation correction model in one embodiment of this application.

[0037] Figure 9 The diagram shown is a schematic block diagram of a PET image self-attenuation correction system according to an embodiment of this application. Detailed Implementation

[0038] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0039] Before providing a further detailed description of the present invention, the nouns and terms used in the embodiments of the present invention are explained, and the nouns and terms used in the embodiments of the present invention are subject to the following interpretations:

[0040] <1> NAC (Non-Attenuation Corrected): No attenuation correction.

[0041] <2> PET (Positron Emission Tomography) is a functional molecular imaging technique based on radioactive tracers. It detects the distribution of tracers in the body to reflect functional activities such as tissue metabolism and blood flow, and is widely used in the clinical diagnosis of tumors, nervous system diseases, and cardiovascular diseases.

[0042] <3> AC (Attenuation Corrected): Attenuation correction.

[0043] <4> SUV (Standardized Uptake Value): This is a core indicator for quantitative analysis of PET images, used to standardize the uptake of radioactive tracers by tissues or lesions in the body.

[0044] <5> CT (Computed Tomography): Computed tomography is an imaging technique that uses X-rays to penetrate the human body and processes the tomographic scan data through a computer to generate clear tomographic images of the internal structure of the human body. Its core advantage is that it can accurately display the density differences of anatomical structures, making up for the shortcomings of ordinary X-ray planar imaging.

[0045] <6> CNN (Convolutional Neural Network): A convolutional neural network is a deep learning model that excels at processing grid-structured data such as images. It extracts local features (such as edges and textures) from images through convolution operations and achieves feature dimensionality reduction and abstraction through operations such as pooling. It is widely used in image recognition, segmentation, generation and other fields.

[0046] <7> GAN (Generative Adversarial Network): A generative adversarial network consisting of two subnetworks, a generator and a discriminator, which generate data through adversarial training between the two.

[0047] <8> DN (Denoising Network): Denoising network.

[0048] <9> U-Net: U-shaped network is a codec-decoder convolutional neural network designed specifically for medical image processing. Its core is adapted to medical imaging tasks such as image segmentation, image generation, and image denoising.

[0049] With the development of artificial intelligence, deep learning has been applied to PET self-attenuation correction (generating AC-PET images from NAC-PET images). Existing methods are mainly divided into two categories: direct attenuation correction methods and indirect attenuation correction methods. Direct attenuation correction methods have a simple process, but due to the lack of anatomical information guidance, the generated AC-PET images are overly smoothed and have limited quantitative accuracy. Indirect attenuation correction methods supplement anatomical information by synthesizing pseudo-CT through a two-step serial process, but they face the challenges of cross-modal synthesis, error accumulation, and optimization isolation. At the same time, the traditional CNN or GAN architectures used in existing methods have shortcomings such as insufficient ability to capture details and susceptibility to modality collapse, making it difficult to match individual differences and meet the high-precision quantitative needs of clinical practice.

[0050] In order to solve the technical problems existing in the background art, this application provides a PET image self-attenuation correction method, system, medium and product, which aims to solve the technical problems of insufficient correction accuracy, inaccurate quantification and error accumulation in the prior art.

[0051] To facilitate understanding of the embodiments of this application, firstly, in conjunction with Figure 1 Detailed explanation. Figure 1 A flowchart illustrating the PET image self-attenuation correction method in this embodiment of the invention is shown. The PET image self-attenuation correction method in this embodiment mainly includes the following steps:

[0052] S1: Construct a PET self-attenuation correction model; the PET self-attenuation correction model includes a two-branch residual diffusion model.

[0053] In this embodiment, the PET self-attenuation correction model is constructed in the following ways:

[0054] S11: Construct a training dataset; the training dataset includes multiple unattenuated corrected PET historical images, attenuated corrected PET historical images corresponding to each of the unattenuated corrected PET historical images, and CT historical images.

[0055] In this embodiment, unattenuated corrected PET historical images, attenuated corrected PET historical images, and CT historical images from different patients are collected to construct a training dataset. Each unattenuated corrected PET historical image corresponds to one attenuated corrected PET historical image and one CT historical image.

[0056] S12: Input each of the unattenuated corrected PET historical images, the attenuated corrected PET historical images corresponding to each of the unattenuated corrected PET historical images, and the CT historical images into the bi-branch residual diffusion model, perform bi-branch learning operation, and construct a primary PET self-attenuation correction model based on the preset total loss function.

[0057] In this embodiment, as Figure 2 The diagram illustrates the flowchart for constructing a primary PET self-attenuation correction model in an embodiment of the present invention. The dual-branch residual diffusion model includes a PET branch residual diffusion model and a CT branch residual diffusion model; wherein, the method for constructing the primary PET self-attenuation correction model includes:

[0058] S121: Input the unattenuated PET historical image and the attenuated PET historical image corresponding to the unattenuated PET historical image into the PET branch residual diffusion model, and perform PET branch learning operation to generate a coarse attenuated PET historical image corresponding to the unattenuated PET historical image.

[0059] S122: Input the unattenuated corrected PET historical image and the corresponding CT historical image into the CT branch residual diffusion model, and perform CT branch learning operation to generate a coarse reconstructed CT historical image corresponding to the unattenuated corrected PET historical image.

[0060] In this embodiment, as Figure 3 The diagram illustrates the training process for coarse attenuation correction in this embodiment of the invention. During the training of the coarse attenuation correction stage, two residual diffusion models (RDMs) are used in parallel: a PET branch residual diffusion model and a CT branch residual diffusion model. The PET branch residual diffusion model learns the mapping relationship between an unattenuated PET image (NAC-PET) and a coarsely attenuated PET image (coarse AC-PET), while the CT branch residual diffusion model learns the mapping relationship between an unattenuated PET image (NAC-PET) and a coarse CT image. This stage aims to quickly generate an intermediate result containing basic functional information and preliminary anatomical contours, laying the foundation for subsequent fine attenuation correction.

[0061] In this embodiment, the PET branch residual diffusion model includes a first PET denoising network (i.e., Figure 3 DN in the denoising network PET The CT branch residual diffusion model includes the first CT denoising network (i.e. Figure 3 DN in the denoising network CT Both the first PET denoising network and the first CT denoising network adopt an encoder-decoder architecture based on attention U-Net. In this stage, real AC-PET images (i.e., attenuated corrected PET historical images corresponding to unattenuated corrected PET historical images) and CT images (i.e., CT historical images corresponding to unattenuated corrected PET historical images) are used as labels for supervised training.

[0062] In this embodiment, in the PET branch residual diffusion model, the attenuated corrected PET historical image (i.e. Figure 3 Starting with the AC-PET image, the model progressively degrades the image into a random distribution with Gaussian noise through multiple steps of adding "residual 1 + Gaussian noise," ultimately reducing it to "NAC-PET + Gaussian noise." Each noise addition step generates an "intermediate variable," and the entire process is a fixed, computable random process. The goal is for the model to learn the distribution pattern of "images changing from clear to noisy." Starting with the forward-diffused final noisy image (NAC-PET + Gaussian noise), the denoising network DN... PET By gradually "eliminating noise," a clear image is restored from the reverse perspective, and a coarse attenuation-corrected PET historical image is output. Figure 3 (coarse AC-PET in the image). It should be noted that residual 1 is the unattenuated corrected PET history image (i.e., Figure 3 NAC-PET images and attenuated corrected PET history images (i.e., NAC-PET images) Figure 3 The pixel difference image between AC-PET).

[0063] In this embodiment, similarly, in the CT branch residual diffusion model, starting with historical CT images, the image is gradually "degraded" into a random distribution with Gaussian noise through multiple steps of "residual 2 + Gaussian noise" superposition, ultimately degrading to "NAC-PET + Gaussian noise". Starting with the final noisy image from forward diffusion (NAC-PET + Gaussian noise), the denoising network DN... CT By gradually "eliminating noise," a clear image is restored from the reverse perspective, and a coarsely reconstructed historical CT image is output. Figure 3 The denoising network learns the mapping relationship of "recovering the clear image from the previous step from the intermediate variable / noisy image". Essentially, through supervised training (using real AC-PET / CT as labels), the model masters the ability of "denoising + generation". It should be noted that residual 2 represents the unattenuated corrected PET historical image (i.e.,...). Figure 3 NAC-PET images and historical CT images (i.e., NAC-PET) Figure 3 The pixel difference image between CT (in the image).

[0064] S123: Based on the coarse attenuation-corrected PET historical image and the coarse reconstructed CT historical image corresponding to the unattenuation-corrected PET historical image, perform feature fusion and feature decoding operations to generate a fine attenuation-corrected PET historical image and a fine CT reconstructed image corresponding to the unattenuation-corrected PET historical image.

[0065] In this embodiment, as Figure 4 The diagram illustrates the training process for fine attenuation correction in an embodiment of the present invention. The PET self-attenuation correction model further includes a dual-branch mutual guidance module; wherein, based on the coarse attenuation-corrected PET historical image and the coarse reconstructed CT historical image corresponding to the unattenuated PET historical image, feature fusion and feature decoding operations are performed to generate the fine attenuation-corrected PET historical image and the fine CT reconstructed image corresponding to the unattenuated PET historical image, including:

[0066] S1231: Based on the dual-branch mutual guidance module, and according to the coarse-attenuated corrected PET historical image and the coarse-reconstructed CT historical image corresponding to the unattenuated corrected PET historical image, a feature fusion operation is performed to generate PET branch fusion historical features and CT branch fusion historical features. The method includes:

[0067] (1) Based on the PET branch residual diffusion model, perform the first denoising feature extraction operation on the coarse attenuation corrected PET historical image corresponding to the unattenuated corrected PET historical image to generate PET branch interaction historical features.

[0068] In this embodiment, the PET branch residual diffusion model further includes a second PET denoising network. The second PET denoising network performs a first denoising feature extraction operation on the coarse attenuation-corrected PET historical image. After convolution and pooling processing, it extracts the metabolic features of the coarse attenuation-corrected PET historical image to generate PET branch interaction historical features f. P The architecture of the second PET denoising network is similar to that of the first PET denoising network, and will not be described in detail here.

[0069] (2) Based on the CT branch residual diffusion model, perform a second denoising feature extraction operation on the coarse reconstructed CT historical image corresponding to the unattenuated corrected PET historical image to generate CT branch interaction historical features.

[0070] In this embodiment, the CT branch residual diffusion model further includes a second CT denoising network. The second CT denoising network performs a second denoising feature extraction operation on the coarsely reconstructed CT history image. After convolution and pooling processing, it extracts human anatomical structure features to generate CT branch interaction history features f. C The architecture of the second CT denoising network is similar to that of the first CT denoising network, and will not be described in detail here.

[0071] (3) Based on the linear layer of the dual-branch mutual guidance module, feature transformation operations are performed on the PET branch interaction history features and CT branch interaction history features respectively to generate PET branch attention history features and CT branch attention history features.

[0072] In this embodiment, as Figure 5 The diagram illustrates the structure of the dual-branch mutual guidance module in an embodiment of the present invention. The dual-branch mutual guidance module includes a linear layer, a multi-head attention layer, and a normalization layer.

[0073] In this embodiment, the PET branch interaction history feature f P CT branch interaction history features f C Feature transformation operations are performed through the linear layer of the dual-branch mutual guidance module to map the query (Q), key (K), and value (V) required by the cross-attention mechanism, and to generate PET branch attention history features and CT branch attention history features.

[0074] (4) Based on the cross-attention layer of the dual-branch mutual guidance module, feature fusion operation is performed on the PET branch attention history features and the CT branch attention history features respectively to generate PET branch fusion history features and CT branch fusion history features.

[0075] In this embodiment, the PET branch attention history features include the PET branch query history features Q. P PET branch bond history characteristics K P and the historical characteristics of PET branch values ​​V P The CT branch attention history features include the CT branch query history features Q. C CT branch bond history features K C and CT branch value historical characteristics V CThe method for generating PET branch fusion historical features and CT branch fusion historical features based on the cross-attention layer of the dual-branch mutual guidance module includes: performing a first multi-head attention calculation operation on the PET branch query historical features, the CT branch key historical features, and the CT branch value historical features based on the cross-attention layer of the dual-branch mutual guidance module; and simultaneously performing a second multi-head attention calculation operation on the CT branch query historical features, the PET branch key historical features, and the PET branch value historical features to generate PET branch fusion historical features and CT branch fusion historical features.

[0076] In this embodiment, the PET branch interaction history feature f P The linear layer of the dual-branch mutual guidance module generates PET branch attention history features, i.e., PET branch query history features Q. P PET branch bond history characteristics K P and the historical characteristics of PET branch values ​​V P Similarly, the CT branch interaction history feature f C The linear layer of the dual-branch mutual guidance module generates CT branch attention history features, i.e., CT branch query history features Q. C CT branch bond history features K C and CT branch value historical characteristics V C .

[0077] In this embodiment, the generated PET branch query historical feature Q P PET branch bond history characteristics K P Historical characteristics of PET branch values ​​V P Historical features of CT branch query Q C CT branch bond history features K C CT branch value historical characteristics V C Mutual guidance is achieved through a cross-attention mechanism, enabling feature exchange and fusion between the PET branch residual diffusion model and the CT branch residual diffusion model in each back-diffusion step. Specifically, the historical features Q of the PET branch query are used... P Historical features of CT branching bond K C CT branch value historical characteristics V C The first multi-head attention computation operation is performed to integrate CT information into AC-PET features, and after processing by a normalization layer (Norm), PET branch fusion history features f are generated. PC Similarly, the CT branch query historical feature Q CHistorical characteristics of PET branched bonds K P Historical characteristics of PET branch values ​​V P A second multi-head attention computation operation is performed to integrate AC-PET information into CT features, and after processing by a normalization layer (Norm), CT branch fusion history features f are generated. CP .

[0078] It is worth noting that, in order to fully utilize the complementary information between PET and CT images and achieve mutual guidance and collaborative optimization, this application introduces a dual-branch mutual guidance module. During the fine attenuation correction stage, this module exchanges and fuses the features extracted from the PET and CT branch residual diffusion models in the back-diffusion step. Specifically, through a cross-attention mechanism, multi-head attention is performed between the query of one branch and the key and value of the other branch to achieve effective feature fusion. This mutual guidance mechanism enables CT and PET to assist each other during the generation process, jointly improving the image generation quality.

[0079] S1232: Based on the PET branch residual diffusion model and according to the PET branch fusion history features, perform a first decoding operation to generate a finely attenuated PET history image corresponding to the unattenuated PET history image.

[0080] In this embodiment, as Figure 6 The diagram illustrates the structure of the lesion perception refinement module in an embodiment of the present invention. The PET self-attenuation correction model further includes the lesion perception refinement module; wherein, based on the PET branch residual diffusion model and according to the PET branch fusion history features, the method of performing a first decoding operation to generate a finely attenuated corrected PET history image corresponding to the unattenuated corrected PET history image includes:

[0081] (1) Based on the lesion perception refinement module, perform feature segmentation operation on the coarse attenuation-corrected PET historical image corresponding to the unattenuation-corrected PET historical image to generate a lesion historical mask.

[0082] (2) Based on the PET branch residual diffusion model and according to the lesion history mask, perform a first decoding operation on the PET branch fusion history features to generate a fine attenuation-corrected PET history image corresponding to the unattenuated PET history image.

[0083] In this embodiment, the feature decoding operation includes a first decoding operation and a second decoding operation. Using the lesion perception refinement module (i.e., the lesion segmentation network), feature segmentation is performed on the coarse-attenuated corrected PET historical image corresponding to the unattenuated corrected PET historical image to generate a lesion history mask. The PET branch fuses historical features f. PC The data is input into the PET branch residual diffusion model and processed by upsampling and normalization operations in the sampling layer and normalization layer. Then, it is input into the decoder of the PET branch residual diffusion model along with the lesion history mask to perform the first decoding operation, thereby generating a fine attenuation-corrected PET history image (fine AC-PET) corresponding to the unattenuated PET history image.

[0084] In this embodiment, to address the issue of insufficient model focus on lesion regions, this application introduces a lesion perception refinement module. First, a pre-trained and frozen lesion segmentation network is used to extract a binarized lesion history mask M from the coarse attenuation-corrected PET history image corresponding to the unattenuated PET history image. L To include the boundary information of the lesions, the lesion history mask, after morphological expansion, is introduced into the decoder of the PET branch residual diffusion model through a spatial adaptive normalization mechanism. This modulates the spatial scale and bias of the feature map, guiding the fusion of historical features f in the PET branch. PC This decoding allows the model to focus more on the lesion area, thereby improving the quantitative fidelity of the lesion area.

[0085] In this embodiment, the lesion segmentation network can adopt a U-shaped convolutional neural network (U-Net). It is first pre-trained based on the attenuated and corrected PET historical image dataset. During training, the attenuated and corrected PET historical images are used as input and the corresponding gold standard lesion mask is used as the supervision label. The features of metabolically active lesion regions are extracted through convolution operation, and the feature dimensionality is reduced through pooling operation. Then, the lesion boundary detail features are passed by deconvolution upsampling and skip connections, and a binary lesion mask is output. After training, all network parameters are frozen to ensure that the parameters are not updated when extracting the lesion mask in the future.

[0086] S1233: Based on the CT branch residual diffusion model and according to the CT branch fusion history features, perform a second decoding operation to generate a fine CT reconstruction image corresponding to the unattenuated corrected PET history image.

[0087] In this embodiment, CT branch fusion historical features f CPThe data is input into the CT branch residual diffusion model and processed by upsampling and normalization operations in the sampling layer and normalization layer. Then, it enters the decoder of the CT branch residual diffusion model and performs the second decoding operation to generate a fine CT reconstruction image (fine CT) corresponding to the unattenuated corrected PET historical image.

[0088] S124: Based on the finely attenuated PET historical image and the finely reconstructed CT image corresponding to the unattenuated PET historical image, the total loss function is calculated, and a primary PET self-attenuation correction model is constructed accordingly.

[0089] In this embodiment, as Figure 7 The diagram illustrates the structure of the attenuation prior module in this embodiment of the invention. Based on the finely attenuated corrected PET historical image corresponding to the unattenuated corrected PET historical image and the finely reconstructed CT image, the total loss function is calculated. The method for constructing the primary PET self-attenuation correction model includes:

[0090] (1) The loss function of the PET branch residual diffusion model is calculated based on the finely attenuated PET historical image corresponding to the unattenuated PET historical image.

[0091] In this embodiment, based on the finely attenuated PET historical image corresponding to the unattenuated PET historical image, the attenuated PET historical image (the real AC-PET image) corresponding to the unattenuated PET historical image is used as a label. The mean squared error loss function (MSE Loss) can be used to calculate the pixel-level numerical difference between the finely attenuated PET historical image and the attenuated PET historical image to obtain the loss function of the PET branch residual diffusion model.

[0092] (2) The loss function of the CT branch residual diffusion model is calculated based on the fine CT reconstructed image corresponding to the unattenuated corrected PET historical image.

[0093] In this embodiment, based on the detailed CT reconstructed image corresponding to the unattenuated corrected PET historical image, the CT historical image (the real CT image) corresponding to the unattenuated corrected PET historical image is used as a label. The mean squared error loss function (MSE Loss) can be used to calculate the pixel-level numerical difference between the detailed CT reconstructed image and the CT historical image to obtain the loss function of the CT branch residual diffusion model.

[0094] (3) Input the unattenuated PET historical image and the fine CT reconstruction image corresponding to the unattenuated PET historical image into the pre-built attenuation prior module to generate an attenuated PET reference image.

[0095] In this embodiment, the unattenuated corrected PET historical image and the corresponding detailed CT reconstructed image are stitched together (i.e., Figure 7 After C), it is input into the attenuation prior module (i.e. Figure 7 In the attenuation prior network, attenuated PET reference images are generated.

[0096] (4) Based on the attenuated PET reference image and the finely attenuated PET historical image corresponding to the unattenuated PET historical image, the attenuation prior constraint loss function is calculated.

[0097] In this embodiment, the mean squared error loss function (MSE Loss) is used to calculate the pixel-level numerical difference between the attenuated PET reference image and the finely attenuated PET historical image, so as to obtain the attenuation prior constraint loss function.

[0098] (5) Based on the loss function of the PET branch residual diffusion model, the loss function of the CT branch residual diffusion model and the attenuation prior constraint loss function, the total loss function is calculated, and the primary PET self-attenuation correction model is constructed accordingly.

[0099] In this embodiment, the formula for calculating the total loss function includes:

[0100] Formula (1)

[0101] Where L represents the total loss function; The loss function represents the PET branch residual diffusion model; The loss function represents the CT branch residual diffusion model; This represents the decaying prior constraint loss function; This indicates the first weight, which is 0.25; This indicates the second weight, which is 0.75.

[0102] In this embodiment, , , The loss function can be calculated based on the mean squared error loss function (MSE Loss), but it can be replaced with other loss function types.

[0103] In this embodiment, the total loss function is calculated according to formula (I), and the parameters of the PET branch residual diffusion model, the CT branch residual diffusion model, and the dual-branch mutual guidance module are adjusted accordingly to construct the primary PET self-attenuation correction model.

[0104] It is worth noting that, to improve the realism of the generated fine CT images and fine attenuation-corrected PET images, this application introduces physical prior constraints, i.e., a pre-trained attenuation prior network. Specifically, during pre-training, the unattenuated PET training image set and the real CT training image set are used as inputs, and the attenuated real PET training image set is used as supervision for training, simulating the real PET attenuation correction process. During the training of the fine attenuation correction stage, the parameters of this attenuation prior network are frozen and included as part of the total loss function, receiving the fine CT reconstructed images (i.e., images generated by the CT branch residual diffusion model) Figure 7 Fine CT images and unattenuated corrected PET historical images (i.e., fine CT images) and unattenuated corrected PET historical images Figure 7 The NAC-PET model is used as input, and the output is compared with the finely attenuated corrected PET history image generated by the PET branch residual diffusion model (i.e., Figure 7 The attenuation prior constraint loss function is calculated in the fine AC-PET to constrain the generation of fine AC-PET and fine CT, so as to make the model generate results that are more consistent with the real attenuation correction law.

[0105] S13: Perform iterative training on the primary PET self-decay correction model to update the total loss function, and adjust and optimize the parameters of the primary PET self-decay correction model according to the updated total loss function to generate the finally converged PET self-decay correction model.

[0106] In this embodiment, as Figure 8 The diagram illustrates the flowchart of constructing the PET self-attenuation correction model in this embodiment of the invention. The constructed training dataset is input into the model in batches. First, after a coarse attenuation correction stage, coarse attenuation-corrected PET historical images (coarse AC-PET) and coarse reconstructed CT historical images (coarse CT) are generated. Then, in the fine attenuation correction stage, a dual-branch mutual guidance module is used to achieve feature exchange and fusion between the PET branch residual diffusion model and the CT branch residual diffusion model. Based on the lesion perception refinement module, fine attenuation-corrected PET historical images (fine AC-PET) and fine CT reconstructed images (fine CT) are generated. Finally, based on the attenuation prior module, after each round of forward propagation, the total loss function is calculated, and an iterative training process is executed. The parameters of the primary PET self-attenuation correction model are adjusted and optimized according to the updated total loss function until the total loss function converges to a preset threshold, and the quantitative accuracy of the fine AC-PET images output by the model meets clinical requirements. At this point, iteration stops, and the finally converged PET self-attenuation correction model is obtained.

[0107] S2: Based on the acquired unattenuated PET real-time image and the PET self-attenuation correction model, generate an attenuated PET target image.

[0108] In this embodiment, after model training, an end-to-end PET self-attenuation correction model is obtained. In practical applications, inputting the unattenuated real-time PET image into the PET self-attenuation correction model directly yields the attenuated target PET image, achieving fully automatic PET self-attenuation correction. Specifically, firstly, in the coarse attenuation correction stage, after the unattenuated real-time PET image is input into the PET self-attenuation correction model, a coarse attenuated real-time PET image and a coarse reconstructed real-time CT image are generated through a bi-branch residual diffusion model. Secondly, in the fine attenuation correction stage, based on the bi-branch mutual guidance module and the lesion perception refinement module, the coarse attenuated real-time PET image and the coarse reconstructed real-time CT image are jointly optimized and finely reconstructed, ultimately generating a high-fidelity attenuated target PET image.

[0109] In this embodiment, the network model used in this invention (such as U-Net) can be replaced by other models without affecting the content of this application.

[0110] It is worth noting that the PET image self-attenuation correction method of this application has the following advantages:

[0111] (1) No additional CT scan required: The method in this application only requires NAC-PET images to complete attenuation correction, which reduces radiation risk and medical costs and avoids radiation exposure and economic burden caused by additional CT scans.

[0112] (2) Improved calibration accuracy and quantitative fidelity: Through the dual-branch mutual guidance module and the lesion perception refinement module, PET images and CT images guide each other and optimize in collaboration during the generation process. Combined with physical prior constraints, the dependence on synthetic CT is reduced, the propagation and amplification of errors are effectively suppressed, and the accuracy and efficiency of attenuation correction are improved, so that the model can generate AC-PET images with clearer anatomical details and more accurate lesion SUV values.

[0113] (3) High-fidelity image generation: The residual diffusion model is used as the backbone network. Compared with CNN or GAN, it can better capture complex data distribution and generate clearer and more detailed images, avoiding problems such as over-smoothing and mode collapse.

[0114] (4) Fully automated end-to-end process, simple and efficient operation: After the model training is completed, end-to-end self-attenuation correction can be achieved. The operation is simple and efficient, without manual intervention, and easy to clinical application and promotion.

[0115] In the embodiments of this application, terms such as "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, "first weight" and "second weight" are only used to distinguish different weights and do not limit their order. Those skilled in the art will understand that terms such as "first" and "second" do not limit the quantity or execution order, and that terms such as "first" and "second" do not necessarily imply that they are different.

[0116] It should be noted that, in the embodiments of this application, the words "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0117] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0118] Figure 9 This is a schematic block diagram of the PET image self-attenuation correction system provided in the embodiments of this application. Figure 9 As shown, the PET image self-attenuation correction system 900 includes:

[0119] The model building module 901 is used to build a PET self-attenuation correction model; the PET self-attenuation correction model includes a two-branch residual diffusion model.

[0120] The correction module 902 is used to generate a PET target image after attenuation correction based on the acquired unattenuated PET real-time image and the PET self-attenuation correction model.

[0121] In this embodiment, the PET self-attenuation correction model is constructed in the following ways:

[0122] (1) Construct a training dataset; the training dataset includes multiple unattenuated corrected PET historical images, attenuated corrected PET historical images corresponding to each of the unattenuated corrected PET historical images, and CT historical images.

[0123] (2) Input each of the unattenuated PET historical images, the attenuated PET historical images corresponding to each of the unattenuated PET historical images, and the CT historical images into the two-branch residual diffusion model, perform the two-branch learning operation, and construct the primary PET self-attenuation correction model based on the preset total loss function.

[0124] (3) Perform iterative training on the primary PET self-decay correction model to update the total loss function, and adjust and optimize the parameters of the primary PET self-decay correction model according to the updated total loss function to generate the finally converged PET self-decay correction model.

[0125] It should be understood that the specific process of each module performing the above-mentioned steps has been described in detail in the above method embodiments, and will not be repeated here for the sake of brevity.

[0126] It should also be understood that the module division in the embodiments of this application is illustrative and only represents a logical functional division; in actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0127] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described above. Figures 1 to 8 The method of any of the embodiments shown.

[0128] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes computer program code. When the computer program code is run on a computer, it causes the computer to execute... Figures 1 to 8 The method of any of the embodiments shown.

[0129] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0130] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0131] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0135] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs, DVDs), or semiconductor media (e.g., solid-state disks, SSDs, etc.).

[0136] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0137] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0138] In summary, this application provides a PET image self-attenuation correction method, system, medium, and product. It employs a two-stage generation strategy from coarse to fine, and utilizes a PET self-attenuation correction technology based on a bi-branch residual diffusion model. Furthermore, through a bi-branch mutual guidance module and a lesion perception refinement module, PET and CT images mutually guide and collaboratively optimize each other during the generation process. Combined with physical prior constraints, this reduces dependence on synthetic CT, effectively suppresses error propagation and amplification, and improves the accuracy, efficiency, and clinical applicability of attenuation correction. Therefore, this application effectively overcomes the various shortcomings of existing technologies and possesses high industrial applicability.

[0139] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for self-attenuation correction of PET images, characterized in that, include: Construct a PET self-attenuation correction model; the PET self-attenuation correction model includes a two-branch residual diffusion model; Based on the acquired unattenuated real-time PET image and the PET self-attenuation correction model, an attenuated target PET image is generated. The PET self-attenuation correction model is constructed in the following ways: Construct a training dataset; the training dataset includes multiple unattenuated corrected PET historical images, attenuated corrected PET historical images corresponding to each of the unattenuated corrected PET historical images, and CT historical images; Each of the unattenuated corrected PET historical images, the attenuated corrected PET historical images corresponding to each of the unattenuated corrected PET historical images, and the CT historical images are input into the bi-branch residual diffusion model, a bi-branch learning operation is performed, and a primary PET self-attenuation correction model is constructed based on a preset total loss function. Iterative training is performed on the primary PET self-attenuation correction model to update the total loss function, and the parameters of the primary PET self-attenuation correction model are adjusted and optimized according to the updated total loss function to generate the finally converged PET self-attenuation correction model.

2. The PET image self-attenuation correction method according to claim 1, characterized in that, The dual-branch residual diffusion model includes a PET branch residual diffusion model and a CT branch residual diffusion model; wherein, the method for constructing the primary PET self-attenuation correction model includes: The unattenuated PET historical image and the attenuated PET historical image corresponding to the unattenuated PET historical image are input into the PET branch residual diffusion model, and the PET branch learning operation is performed to generate the coarse attenuated PET historical image corresponding to the unattenuated PET historical image. The unattenuated PET historical image and the corresponding CT historical image are input into the CT branch residual diffusion model, and the CT branch learning operation is performed to generate the coarse reconstructed CT historical image corresponding to the unattenuated PET historical image. Based on the coarse attenuation-corrected PET historical image and the coarse reconstructed CT historical image corresponding to the unattenuation-corrected PET historical image, feature fusion and feature decoding operations are performed to generate a fine attenuation-corrected PET historical image and a fine CT reconstructed image corresponding to the unattenuation-corrected PET historical image. Based on the finely attenuated PET historical image and the finely reconstructed CT image corresponding to the unattenuated PET historical image, the total loss function is calculated, and a primary PET self-attenuation correction model is constructed accordingly.

3. The PET image self-attenuation correction method according to claim 2, characterized in that, The PET self-attenuation correction model further includes a dual-branch mutual guidance module; wherein, based on the coarse attenuation-corrected PET historical image and the coarse reconstructed CT historical image corresponding to the unattenuated PET historical image, the method of performing feature fusion and feature decoding operations to generate the fine attenuation-corrected PET historical image and the fine CT reconstructed image corresponding to the unattenuated PET historical image includes: Based on the dual-branch mutual guidance module, and according to the coarse attenuation-corrected PET historical image and the coarse reconstructed CT historical image corresponding to the unattenuated PET historical image, a feature fusion operation is performed to generate PET branch fusion historical features and CT branch fusion historical features. Based on the PET branch residual diffusion model and according to the PET branch fusion history features, a first decoding operation is performed to generate a fine attenuation-corrected PET history image corresponding to the unattenuated PET history image. Based on the CT branch residual diffusion model and according to the CT branch fusion history features, a second decoding operation is performed to generate a fine CT reconstruction image corresponding to the unattenuated corrected PET history image.

4. The PET image self-attenuation correction method according to claim 3, characterized in that, Based on the dual-branch mutual guidance module, and according to the coarse-attenuated corrected PET historical image and the coarse-reconstructed CT historical image corresponding to the unattenuated corrected PET historical image, a feature fusion operation is performed to generate PET branch fusion historical features and CT branch fusion historical features, including the following methods: Based on the PET branch residual diffusion model, a first denoising feature extraction operation is performed on the coarse attenuation-corrected PET historical image corresponding to the unattenuated PET historical image to generate PET branch interaction historical features. Based on the CT branch residual diffusion model, a second denoising feature extraction operation is performed on the coarsely reconstructed CT historical image corresponding to the unattenuated corrected PET historical image to generate CT branch interaction historical features. Based on the linear layer of the dual-branch mutual guidance module, feature transformation operations are performed on the PET branch interaction history features and CT branch interaction history features respectively to generate PET branch attention history features and CT branch attention history features. Based on the cross-attention layer of the dual-branch mutual guidance module, feature fusion operations are performed on the PET branch attention history features and the CT branch attention history features respectively to generate PET branch fusion history features and CT branch fusion history features.

5. The PET image self-attenuation correction method according to claim 4, characterized in that, The PET branch attention history features include PET branch query history features, PET branch key history features, and PET branch value history features; the CT branch attention history features include CT branch query history features, CT branch key history features, and CT branch value history features; wherein, based on the cross-attention layer of the dual-branch mutual guidance module, feature fusion operations are performed on the PET branch attention history features and the CT branch attention history features respectively to generate PET branch fusion history features and CT branch fusion history features, including the following methods: Based on the cross-attention layer of the dual-branch mutual guidance module, a first multi-head attention calculation operation is performed on the PET branch query history features, the CT branch key history features, and the CT branch value history features. At the same time, a second multi-head attention calculation operation is performed on the CT branch query history features, the PET branch key history features, and the PET branch value history features to generate PET branch fusion history features and CT branch fusion history features.

6. The PET image self-attenuation correction method according to claim 3, characterized in that, The PET self-attenuation correction model further includes a lesion perception refinement module; wherein, based on the PET branch residual diffusion model and according to the PET branch fusion history features, the first decoding operation is performed to generate a finely attenuated corrected PET history image corresponding to the unattenuated corrected PET history image, including the following methods: Based on the lesion perception refinement module, feature segmentation is performed on the coarse attenuation-corrected PET historical image corresponding to the unattenuation-corrected PET historical image to generate a lesion historical mask. Based on the PET branch residual diffusion model and according to the lesion history mask, a first decoding operation is performed on the PET branch fusion history features to generate a fine attenuation-corrected PET history image corresponding to the unattenuated PET history image.

7. The PET image self-attenuation correction method according to claim 2, characterized in that, Based on the finely attenuated PET historical image corresponding to the unattenuated PET historical image and the fine CT reconstructed image, the total loss function is calculated, and the methods for constructing the primary PET self-attenuation correction model include: Based on the finely attenuated corrected PET historical image corresponding to the unattenuated corrected PET historical image, the loss function of the PET branch residual diffusion model is calculated. The loss function of the CT branch residual diffusion model is calculated based on the detailed CT reconstructed image corresponding to the unattenuated corrected PET historical image. The unattenuated PET historical image and the corresponding fine CT reconstructed image are input into a pre-built attenuation prior module to generate an attenuated PET reference image. Based on the attenuated and corrected PET reference image, and according to the finely attenuated and corrected PET historical image corresponding to the unattenuated and corrected PET historical image, the attenuation prior constraint loss function is calculated. Based on the loss functions of the PET branch residual diffusion model, the CT branch residual diffusion model, and the attenuation prior constraint loss function, the total loss function is calculated, and a primary PET self-attenuation correction model is constructed accordingly.

8. A PET image self-attenuation correction system, characterized in that, include: The model building module is used to build a PET self-attenuation correction model; the PET self-attenuation correction model includes a two-branch residual diffusion model. The correction module is used to generate a PET target image after attenuation correction based on the acquired uncorrected real-time PET image and the PET self-attenuation correction model. The PET self-attenuation correction model is constructed in the following ways: Construct a training dataset; the training dataset includes multiple unattenuated corrected PET historical images, attenuated corrected PET historical images corresponding to each of the unattenuated corrected PET historical images, and CT historical images; Each of the unattenuated corrected PET historical images, the attenuated corrected PET historical images corresponding to each of the unattenuated corrected PET historical images, and the CT historical images are input into the bi-branch residual diffusion model, a bi-branch learning operation is performed, and a primary PET self-attenuation correction model is constructed based on a preset total loss function. Iterative training is performed on the primary PET self-attenuation correction model to update the total loss function, and the parameters of the primary PET self-attenuation correction model are adjusted and optimized according to the updated total loss function to generate the finally converged PET self-attenuation correction model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the PET image self-attenuation correction method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes computer program code that, when run on a computer, causes the computer to implement the PET image self-attenuation correction method as described in any one of claims 1 to 7.

Images