A low-dose pet / mr denoising method based on individualized brain partitioning

Abstract

This invention discloses a low-dose PET / MR denoising method based on individualized brain region partitioning. The method includes: acquiring a low-dose PET image and a corresponding registered T1-weighted MR image for a target brain region; and obtaining a denoised reconstructed PET image based on the low-dose PET image and the MR image using a trained encoder-decoder model. The encoder-decoder model includes an attention branch and a visual Mamba branch, and incorporates the adjacency relationships of static brain structures with neuroanatomical significance into the image feature modeling process. The attention branch is used to model the shape similarity of local metabolic patterns and calculates attention based on the square of the Pearson correlation coefficient, while the visual Mamba branch is used to model global long-range dependencies. This invention, by synergistically combining local structure modeling and long-range dependency modeling, enables the model to restore detailed texture while taking into account overall metabolic distribution characteristics, thereby improving image visual quality while maintaining quantitative consistency at the brain region level.

Description

Technical Field

[0001] This invention relates to the field of positron emission tomography (PET) technology in the medical and industrial fields, and more specifically, to a low-dose PET / MR denoising method based on individualized brain partitioning. Background Technology

[0002] Positron emission tomography (PET), as an important molecular imaging technique, has significant applications in early screening of neurodegenerative diseases, brain function research, and efficacy evaluation, especially in detecting metabolic abnormalities in brain diseases such as Alzheimer's disease, where it can provide crucial functional information. The quality of PET images is closely related to the injection dose of the radioactive tracer. While high-dose imaging can achieve a higher signal-to-noise ratio and clearer structural information, it also brings higher radiation risks and cost burdens. Therefore, reducing the PET scan dose while ensuring safety and reproducibility has become an important development direction in current clinical and research fields. However, low-dose PET images often suffer from significant statistical noise, decreased contrast, and degradation of structural details, affecting not only the visual quality of the images but also interfering with quantitative indicators such as the standardized uptake value (SUV) of key brain regions, thus limiting its application in accurate diagnosis and long-term follow-up.

[0003] PET imaging suffers from issues such as photon attenuation, scattering, and statistical noise. Especially with the reduction of tracer injection doses to minimize radiation risks, low-dose PET images often exhibit decreased signal-to-noise ratio, structural blurring, and increased quantitative errors, thus affecting the reliability of clinical diagnosis. To address these issues, current mainstream techniques typically rely on PET / CT or PET / MR systems. These systems utilize anatomical information from CT or MR images for attenuation correction and image enhancement, while incorporating deep learning methods for image reconstruction and denoising. In recent years, with the development of artificial intelligence, deep learning-based CT-free reconstruction, cross-modal generation, and end-to-end image restoration methods have become research hotspots. These methods improve image quality and reduce dependence on additional modalities through data-driven approaches, representing a significant development trend in PET imaging technology.

[0004] In the prior art, patent application CN111436958A provides a CT image generation method for attenuation correction of PET images. This scheme uses uncorrected PET images as input data, and constructs a deep learning model to extract and map features from the PET images, generating corresponding pseudo-CT images for subsequent attenuation correction. The research "Learning CT-free attenuation-corrected total-body PET images through deep learning," published in European Radiology in 2024 by Wenbo Li et al., aims to eliminate the dependence on CT in traditional PET / CT systems and directly generate high-quality attenuation-corrected images from PET data using deep learning methods. Its core idea lies in constructing a model framework based on generative adversarial networks (Cycle-GAN) to map uncorrected PET images to results close to real attenuation-corrected PET images. Simultaneously, tissue structure information is introduced into the model as a priori constraints to enhance the contrast and structural consistency of different tissue regions. This method trains on multiple whole-body PET data to learn the mapping relationship between PET images and corrected images, and uses statistical indicators to quantitatively evaluate the generated results. Experimental results show that the generated images can approach the reference standard in terms of visual quality and quantitative consistency.

[0005] Analysis reveals that existing technologies have the following main drawbacks: (1) Existing methods generally rely on the features of PET images themselves for recovery, lacking explicit use of individualized brain anatomical information of subjects, which easily leads to the denoising results being inconsistent with the real brain structure.

[0006] (2) The dot product or cosine similarity modeling commonly used in the prior art is sensitive to signal intensity translation and scaling, and is prone to weakening the association of semantically consistent regions under low-dose noise.

[0007] (3) Current methods often have the problem of not being able to balance local detail restoration and global structural modeling, which can easily lead to blurred boundaries, over-smoothing or structural breakage. (4) Existing solutions mostly use pixel space for direct modeling, which makes it difficult to fully express the spatial adjacency relationship and anatomical topological constraints between brain regions, resulting in insufficient utilization of structural information.

[0008] (5) Existing models have limited generalization ability in complex noise backgrounds and lack output stability when faced with different subjects, different tracers or different scanning conditions.

[0009] In summary, existing technologies for image denoising primarily rely on deep learning methods, including convolutional neural network-based reconstruction methods, self-attention-based global modeling methods, and state-space models proposed in recent years. While these methods can improve the signal-to-noise ratio and visual quality of images to some extent, they still have several shortcomings. Firstly, most methods rely solely on the image itself for feature learning, lacking explicit modeling of the individualized brain anatomy of the subjects, making it difficult to guarantee the anatomical consistency of the denoising results in spatial structure. Secondly, traditional attention mechanisms typically construct feature associations based on dot product or cosine similarity, and their similarity metrics are highly sensitive to signal intensity translation and scaling. In low-dose noise environments, they are easily affected by intensity fluctuations, leading to weakened associations between semantically similar regions and even erroneous amplification of noisy regions. Furthermore, existing methods struggle to achieve a balance between local detail restoration and global structural modeling, easily resulting in over-smoothing leading to detail loss or structural discontinuities caused by noise propagation. These issues further affect the accuracy and stability of metabolic quantification at the brain region level. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a low-dose PET / MR denoising method based on individualized brain region mapping. This method includes the following steps: Low-dose PET images and corresponding registered T1-weighted MR images were acquired for the target brain region. Based on the low-dose PET images and the MR images, a denoised reconstructed PET image is obtained using a trained encoder-decoder model. The encoder-decoder model includes an attention branch and a visual mamba branch, and introduces the adjacency relationships of static brain structures with neuroanatomical significance into the image feature modeling process to achieve anatomically guided graph feature propagation. The attention branch is used to model the shape similarity of local metabolic patterns and calculate attention based on the square of the Pearson correlation coefficient, while the visual mamba branch is used to model global long-range dependencies.

[0011] Compared with existing technologies, the advantages of this invention are that it provides a denoising and quantification preservation method for low-dose PET / MR brain imaging. It introduces individualized anatomical structural constraints with clear biological significance during image restoration and constructs a feature association mechanism robust to intensity perturbations, reducing noise interference with semantic modeling and enhancing the consistent expression of cross-regional metabolic patterns. This improves image visual quality while maintaining the authenticity and stability of metabolic distribution in key brain regions. In summary, this invention achieves image restoration under anatomical constraints by introducing individualized brain region structural priors; achieves stable cross-regional association expression by constructing a more robust similarity modeling mechanism to amplitude perturbations; preserves local texture details and global anatomical continuity through a collaborative modeling mechanism; achieves structured semantic expression by mapping image features to brain region map space; and improves the adaptability and consistency of the model by introducing static individualized structural priors and robust feature modeling.

[0012] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0014] Figure 1 This is a flowchart of a low-dose PET / MR denoising method based on individualized brain regions according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall process of a low-dose PET / MR denoising method based on individualized brain regions according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a low-dose PET image before noise reduction according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a low-dose PET image after noise reduction according to an embodiment of the present invention; Figure 5 This is a schematic diagram of full-dose PET imaging according to an embodiment of the present invention; In the attached diagram: BC-FFN - Block Convolutional Feedforward Network; DW-Conv - Depthwise Separable Convolution; GELU - Gaussian Error Linear Unit; Silu - Swish Activation Function. Detailed Implementation

[0015] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0016] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0017] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0018] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0019] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0020] To mitigate image noise interference, structural distortion, and metabolic quantification bias under low-dose PET / MR brain imaging conditions, this invention proposes an image denoising method based on individualized brain region anatomical priors and SPCC-Mamba co-modeling. This method uses brain region segmentation results from T1-weighted MR images of subjects as a foundation, constructs static brain structural adjacency relationships with neuroanatomical significance, and incorporates them into the image feature modeling process. Within the encoding and decoding framework, it combines anatomically guided graph structure propagation with visual Mamba (state-space model) feature extraction to achieve a joint expression of brain region spatial relationships and global contextual information. Simultaneously, a similarity modeling mechanism based on the squared Pearson correlation coefficient is introduced to enhance robustness to intensity perturbations and maintain the consistency of local metabolic patterns. This invention improves the quality of denoised images while effectively maintaining the structural integrity and SUV quantitative characteristics of key brain regions.

[0021] Combination Figure 1 and Figure 2 As shown, the provided low-dose PET / MR denoising method based on individualized brain regions includes the following steps: Step S110: Construct an encoding and decoding model architecture. This model includes a SPCC attention branch for modeling shape similarity of local metabolic patterns and a visual Mamba branch for modeling global long-range dependencies. Static brain structure adjacency relationships with neuroanatomical significance are introduced into the image feature modeling process.

[0022] See Figure 2As shown, the encoding / decoding model architecture mainly includes the AG-VIG module (Anatomy-Guided VisionGraph), the efficient SPCC-Mamba module (Squared Pearson Correlation Coefficient-Mamba), and the efficient SPCC (Squared Pearson Correlation Coefficient) attention mechanism module. Figure 2 The diagram illustrates the overall architecture, the structure of SPCC-Mamba, and the structure of the SPCC attention mechanism module.

[0023] The AG-VIG module is the basic feature extraction module. Its core function is to perform preliminary feature processing on brain images. For example, upon receiving brain image data, it first segments the brain to separate brain tissue regions and reduce irrelevant background interference. Then, it transforms the segmented image into a graph structure, captures the topological relationships of brain regions through an adjacency matrix, and extracts graph structure features through 1D convolution (Conv1D), linear transformation (Linear), and ReLU activation function. The Dropout layer is used to prevent model overfitting. Feature fusion: The processed graph features are fused with the original image features to provide richer basic features for subsequent modules.

[0024] The SPCC-Mamba module combines the advantages of the state-space model (Mamba) and correlation calculation to efficiently process brain image features. For example, after splitting the input features, one part enters the SPCC attention branch, where features are extracted using BC-FFN (which includes 3x3 convolution, GELU activation, etc.), while the other part undergoes feature modulation through normalization, linear transformation, and a sigmoid gate mechanism. The features processed by the two branches are concatenated and summed to output a higher-level feature representation, while residual connections preserve the original feature information, enhancing feature expressive power.

[0025] The SPCC attention mechanism module is used to accurately capture the correlation between features in brain images. This invention combines Mamba and SPCC Attention, utilizing Mamba to efficiently process sequential features and the attention mechanism to enhance the correlation between features, thereby further improving the model's ability to perceive key regions in brain images.

[0026] Step S120: Construct a training set and train an encoding / decoding model to achieve denoising of low-dose PET images.

[0027] Step S1: Image acquisition and preprocessing.

[0028] To construct the training set, low-dose PET images and their corresponding T1-weighted MR images were first acquired. The two types of images underwent size unification, intensity normalization, and spatial alignment to ensure consistency in subsequent brain region segmentation and feature fusion. Let the low-dose PET input be denoted as... MR input is The preprocessed result can be expressed as: (1) (2) in, and These represent the normalization and registration preprocessing operators for PET and MR, respectively, used to provide a consistent input basis for subsequent graph modeling. This represents the preprocessed PET image. This represents the preprocessed MR image.

[0029] The constructed training set reflects the correspondence between low-dose PET image samples and standard-dose images.

[0030] Step S2: Perform individualized brain partitioning based on T1-weighted MR images and construct a structural adjacency matrix.

[0031] For example, based on T1-weighted MR images, 189 standard anatomical brain regions were obtained using FreeSurfer (an open-source software suite for analyzing and visualizing neuroimaging data of human brain structure and function), and the geometric centroid of the mask was segmented for each brain region. As the node location. For any node i, calculate its Euclidean distance to other nodes and select the K nearest neighbors to form a spatial adjacency relationship: (3) in, Let i represent the set of the K nearest neighbors of node i. This represents the geometric centroid of the brain region segmentation mask corresponding to node i. This represents the geometric centroid of the brain region segmentation mask corresponding to node j.

[0032] The corresponding individual structural connection weights are defined as follows: (4) in, Let represent the connection strength between node i and node j for the m-th subject. Then, the adjacency matrix of all subjects is averaged, and then symmetricized, self-loop eliminated, and Top-N sparsified to obtain a fixed individualized brain structure adjacency matrix A: (5) Where M represents the total number of subjects, and A is frozen during the training and inference phases to provide static anatomical priors. It is the average of the adjacency matrix of all subjects. This represents the adjacency matrix of the m-th subject. This means taking the N largest elements, for example, N=30.

[0033] Step S3: Anatomy-guided visual map feature mapping and propagation.

[0034] Input feature tensor (Here, X refers to...) After being flattened into a spatial sequence, the image features are mapped to the brain region node space through a learnable projection, forming node features H. Subsequently, two layers of residual graph message propagation are performed on a fixed adjacency matrix A, updating the image features under the topological constraints of the brain regions. The graph propagation operator can be expressed as: (6) (7) in, Let H be a function relating node features H and adjacency matrix A, and D be the degree matrix corresponding to A. These are learnable parameters, used to describe the transformations of a node's own features and its neighbor aggregation features, respectively. It is the first The node feature matrix of the layer, yes The node feature matrix of the layer, It is an index of layers, and more layers can be used. This step constrains image features by the spatial adjacency relationships of brain regions, which helps maintain structural continuity and anatomical rationality.

[0035] Step S4: Inverse projection of graph features and fusion of residuals.

[0036] After graph space propagation is completed, node features are remapped back to the image feature space and residual fusion is performed with the original input to restore spatial resolution and preserve original texture information. This process can be represented as: (8) (9) in, Indicates inverse projection mapping, It was through The result after the calculation To illustrate the enhanced feature residuals, The output features after fusion Indicates will according to B represents the batch size, C represents the number of channels, H represents the height, and W represents the width. This step is used to re-inject anatomical constraints from the graph space into the image domain, allowing subsequent denoising processes to preserve both local details and global structure.

[0037] Step S5: Model using efficient SPCC attention branches.

[0038] The intermediate features are divided into two parts according to the channel dimension. One part is input into an efficient SPCC attention branch to model the shape similarity of local metabolic patterns. For any two local neighborhood vectors Its squared Pearson correlation coefficient can be written as: (10) in, and They represent the vector mean, It is a vector of all 1s. Further construct the kernel function: (11) in, This is the temperature parameter. Based on this kernel function, the attention output can be written as: (12) in, These are query, key, and value matrices, respectively. This corresponds to the feature map. This branch is used to extract stable semantic relevance from local ingestion patterns and reduce the interference of intensity perturbations on feature associations.

[0039] Step S6: Visual Mamba branch modeling, dual-branch fusion, and training using the set objectives.

[0040] Another portion of the features are input into the visual Mamba (state-space model) branch, which models the global long-range dependency through linear projection, depthwise separable convolution, and selective state-space modules. This model is then combined with the SPCC branch results through channel rearrangement and concatenation to obtain the final denoised output. This fusion result can be denoted as: (13) (14) (14) in, and These are the two features after channel splitting. and These represent the SPCC attention branch and the visual Mamba branch, respectively. This step represents the decoder. It allows the model to maintain local structural details while ensuring global metabolic continuity and quantitative consistency.

[0041] Furthermore, by training the encoder-decoder model using the training set until the set optimization objective is met, the optimized model parameters can be obtained.

[0042] In summary, this invention utilizes registered low-dose PET and T1-weighted MR images to construct individualized brain structure maps, and combines anatomically guided graph propagation and efficient SPCC-Mamba feature modeling within an encoder-decoder model framework to achieve unified modeling for image denoising and quantitative preservation of brain regions.

[0043] Step S130: For the target brain region, a trained encoding / decoding model is used to denoise the low-dose PET image to obtain a reconstructed image.

[0044] Once the encoding-decoding model is trained, it can be applied to actual low-dose PET image reconstruction. For example, the model application process is as follows: acquire low-dose PET images of the target brain region and corresponding registered T1-weighted MR images; perform individualized brain partitioning based on the T1-weighted MR images and construct a structural adjacency matrix; then use the trained encoding-decoding model to obtain denoised reconstructed PET images. The model application process is basically similar to the training process and will not be described in detail here.

[0045] It should be understood that the model training process involved in this invention can be performed offline on a server or in the cloud. The trained model can be embedded into an electronic device to achieve real-time PET image reconstruction. This electronic device can be a terminal device or a server. Terminal devices include any terminal device such as mobile phones, tablets, personal digital assistants (PDAs), point-of-sale (POS) terminals, in-vehicle computers, and smart wearable devices. Servers include, but are not limited to, application servers or web servers, and can be independent servers, cluster servers, or cloud servers.

[0046] To further verify the effectiveness of the present invention, experimental verification was conducted. See [link to relevant documentation]. Figures 3 to 5 As shown, where Figure 3 This is a low-dose PET image before noise reduction. Figure 4 It is a low-dose PET image after noise reduction. Figure 5 This is full-dose PET imaging. Experimental results show that the denoised PET images obtained using this invention can maintain the consistency of spatial distribution of brain regions while restoring image quality.

[0047] In summary, compared with the prior art, the present invention has the following main advantages: (1) This invention introduces individualized brain region anatomical priors in the low-dose PET / MR brain imaging denoising task, constructs static brain structure constraint relationships, and realizes anatomical guidance and structure preservation for the image restoration process. By introducing individualized brain region structure priors based on the subject's MR images, anatomical constraints are realized for the low-dose PET denoising process, thereby maintaining the consistency of brain region spatial distribution while restoring image quality, which is beneficial to improving the reliability of structural expression.

[0048] (2) This invention designs an anatomically guided visual image fusion mechanism that combines brain region map propagation with image feature reconstruction to achieve effective injection of spatial correlation information of brain regions and global structural expression. Through this fusion modeling mechanism based on brain region maps and image features, image features can be propagated and updated under the constraints of anatomical topological relationships, effectively enhancing cross-regional information interaction capabilities, thereby improving the integrity and continuity of global structural expression.

[0049] (3) In the feature modeling process, the present invention introduces a similarity measurement mechanism that is insensitive to intensity translation and scaling, which reduces the impact of low-dose noise and intensity perturbation on feature association, thereby improving the stability and robustness of the model under complex imaging conditions.

[0050] (4) The present invention designs the SPCC-Mamba collaborative modeling module, which integrates amplitude-invariant similarity measurement with long-range dependency modeling, thereby improving robust characterization and SUV quantitative preservation under low-dose noise conditions. By synergistically combining local structure modeling and long-range dependency modeling, the model can restore detailed texture while taking into account the overall metabolic distribution characteristics, thereby improving the visual quality of the image while maintaining quantitative consistency at the brain region level.

[0051] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0052] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0053] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0054] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0055] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0056] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A low-dose PET / MR denoising method based on individualized brain regions, comprising the following steps: Low-dose PET images and corresponding registered T1-weighted MR images were acquired for the target brain region. Based on the low-dose PET images and the MR images, a denoised reconstructed PET image is obtained using a trained encoder-decoder model. The encoder-decoder model includes an attention branch and a visual mamba branch, and introduces the adjacency relationships of static brain structures with neuroanatomical significance into the image feature modeling process to achieve anatomically guided graph feature propagation. The attention branch is used to model the shape similarity of local metabolic patterns and calculate attention based on the square of the Pearson correlation coefficient, while the visual mamba branch is used to model global long-range dependencies.

2. The method according to claim 1, characterized in that, The anatomically guided graph feature propagation is achieved according to the following steps: After flattening the input feature tensor into a spatial sequence, it is mapped onto the brain region node space through a learnable projection to form node features H. Residual graph message propagation is performed on the individualized brain structure adjacency matrix A, updating image features under brain region topological constraints. The graph propagation operator is represented as: in, Let H be a function relating node features H and adjacency matrix A, and D be the degree matrix corresponding to A. For learnable parameters, It is the first The node feature matrix of the layer, yes The node feature matrix of the layer, It is the index of the layer.

3. The method according to claim 2, characterized in that, The individualized brain structure adjacency matrix A is obtained according to the following steps: Multiple standard anatomical brain regions were obtained based on T1-weighted MR images, and the geometric centroid of the mask was segmented for each brain region. As the node location, calculate the Euclidean distance between node i and other nodes, and select the K nearest neighbors to form a spatial adjacency relationship, represented as: in, Let i represent the set of the K nearest neighbors of node i. This represents the geometric centroid of the brain region segmentation mask corresponding to node i. This represents the geometric centroid of the brain region segmentation mask corresponding to node j; The corresponding individual structural connection weights are defined as follows: in, This represents the connection strength between the m-th subject and node i and node j; The adjacency matrix of all subjects was averaged, and then symmetrization, self-loop elimination, and Top-30 sparsification were performed to obtain the individualized brain structure adjacency matrix A: Where M is the total number of subjects, It is the average of the adjacency matrix of all subjects. This represents the adjacency matrix of the m-th subject. This means taking the N largest elements.

4. The method according to claim 3, characterized in that, After graph feature propagation, the node features are remapped back to the image feature space and residually fused with the original input to restore spatial resolution and preserve the original texture information, as shown below: in, Indicates inverse projection mapping, It was through The result after the calculation The feature residuals after the graph-guided enhancement. The output features after fusion Indicates will according to B represents batch size, C represents number of channels, H represents height, and W represents width.

5. The method according to claim 1, characterized in that, The attention branch calculates attention according to the following steps: For any two local neighborhood vectors Its squared Pearson correlation coefficient is expressed as: in, and They represent the vector mean, It is a vector consisting entirely of 1s; Construct a kernel function, expressed as: in, Temperature parameters; Based on this kernel function, the attention is calculated, expressed as: in, These are query, key, and value matrices, respectively. This is the corresponding feature mapping.

6. The method according to claim 1, characterized in that, The reconstructed PET image is represented as follows: in: in, and These are the two features after channel splitting. and These represent the SPCC attention branch and the visual Mamba branch, respectively. This indicates the decoder.

7. The method according to claim 1, characterized in that, The attention branch is constructed using a block-type convolutional feedforward network BC-FFN, and the attention mechanism is calculated using the squared Pearson correlation coefficient.

8. The method according to claim 1, characterized in that, The visual Mamba branch comprises a linear projection layer, a depth-separable convolutional layer, and a selective state space module.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer device comprising a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images