A low-dose nuclear medicine PET image reconstruction method and system

By using a low-dose PET reconstruction model based on a U-shaped codec architecture, frequency domain decoupling and Hilbert curve scanning strategies, combined with a full-dose semantic awareness cueing branch, the image degradation problem caused by photon starvation in low-dose PET imaging was solved, and high-quality image reconstruction was achieved.

CN123118799APending Publication Date: 2026-11-17WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611384985.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing low-dose PET imaging techniques suffer from photon starvation during reconstruction, resulting in low signal-to-noise ratio, statistical noise, and artifacts, making it difficult to accurately reconstruct high-quality images, especially with the loss of global structural integrity and fine structural details.

Method used

A low-dose PET reconstruction model based on a U-shaped codec architecture is adopted. By decomposing the features into low-frequency and high-frequency components through frequency domain decoupling, the model is constructed using a global state space model and a Hilbert curve scanning strategy. Combined with full-dose semantic awareness prompts for branching and feature distribution alignment, the reconstruction quality is improved.

Benefits of technology

It significantly improves the structural fidelity and quantitative accuracy of low-dose PET images, effectively restores global structure and local details, and enhances the overall quality of image reconstruction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118799A_ABST
    Figure CN123118799A_ABST
Patent Text Reader

Abstract

The application provides a low-dose nuclear medicine PET image reconstruction method and system, comprising: acquiring a low-dose nuclear medicine PET image and a corresponding full-dose PET image, and constructing a training data set; building a low-dose PET reconstruction model based on a U-shaped coding and decoding architecture; inputting the low-dose PET image into the reconstruction model, supervising and training the model through a combined loss function, and obtaining a trained reconstruction model; inputting a low-dose PET image to be reconstructed into the trained reconstruction model, and obtaining a reconstructed PET image. The application can reduce patient radiation exposure and improve the safety of clinical examination while reconstructing high-quality PET images. As an initial step for downstream nuclear medicine tasks, the application can reduce the clinical PET image acquisition time and improve the medical experience, is suitable for precise quantitative analysis and auxiliary diagnosis of major diseases such as nerves, tumors and hearts, and is widely applied to the fields of nuclear medicine image processing and intelligent diagnosis and treatment technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for reconstructing low-dose nuclear medicine PET images. Background Technology

[0002] Positron emission tomography (PET) is a molecular imaging technique that enables non-invasive in vivo visualization and quantitative analysis of physiological and metabolic processes. This technique relies on the administration of radiolabeled tracers, such as 18F-fluorodeoxyglucose (FD). 18 FFDG) and 68Ga-labeled FAP inhibitors ( 68 Tracers (Ga-FAPI) accumulate in metabolically active tissues (such as tumors) by increasing glucose uptake. Upon radioactive decay, the emitted positrons annihilate with electrons, producing a pair of simultaneously occurring gamma photons, which are detected by a scanner. By accumulating millions of such detection events, the three-dimensional distribution of the tracer can be reconstructed, generating images that reflect regional metabolic activity with high sensitivity and quantitative accuracy. PET is widely used in the diagnosis and treatment of cancer, neurodegenerative diseases, and cardiovascular diseases. However, its clinical value is limited by concerns about radiation exposure, which is associated with potential risks, including carcinogenicity, genetic damage, and deterministic effects on radiation-sensitive organs, especially in children, pregnant women, and patients requiring repeated scans. Reducing the tracer dose or shortening the scan time—i.e., low-dose PET imaging—has become an ideal but challenging approach.

[0003] However, reduced dose directly leads to a decrease in the number of detected coincidence events, a condition commonly referred to as photon starvation. Under photon starvation conditions, limited photon counts manifest as low signal-to-noise ratio (SNR), significant statistical noise, and severe artifacts, with uneven degradation across different frequency bands. Specifically, low-frequency components that determine the overall structure are more severely affected, while high-frequency components encoding fine details are relatively less affected. Therefore, loss of both overall structural integrity and fine structural detail is inevitable. In this context, low-dose PET refers to imaging using a dose lower than the standard dose level, typically corresponding to a DRF greater than 1. These degradations collectively impair quantitative accuracy, such as the standardized uptake value (SUV), and diagnostic reliability, making the reconstruction of high-quality images from low-dose PET data an extremely challenging task.

[0004] To address these challenges, traditional reconstruction methods rely on manually designed prior information to suppress noise in low-dose PET images. However, accurately reconstructing full-dose images remains challenging due to insufficient photon counts, as these methods often produce overly smoothed results that lose structural details, potentially leading to missed lesions or inaccurate tumor delineation.

[0005] With the rapid development of deep learning, data-driven methods have made significant progress in learning the nonlinear mapping between low-dose and full-dose PET images. U-Net and its 3D extensions have demonstrated superior reconstruction performance. However, Convolutional Neural Networks (CNNs) are inherently limited by their local receptive fields, restricting their ability to model long-range spatial dependencies crucial to global structural integrity. To overcome this limitation, Transformer (ViT)-based methods have been introduced into the low-dose PET reconstruction domain to enhance global modeling capabilities. Nevertheless, their quadratic computational complexity has prompted efforts to design more efficient architectures. Furthermore, Generative Adversarial Networks (GANs) have been used to further improve reconstruction quality through adversarial learning.

[0006] Despite significant progress in existing methods, most remain based on two-dimensional (layer-by-layer) or 2.5D modeling paradigms, failing to fully utilize the inherent three-dimensional characteristics of medical images and thus limiting their performance in practical applications. However, this reliance on layer-by-layer modeling introduces inherent limitations. The lack of explicit inter-layer continuity modeling often results in noticeable axial artifacts in the reconstructed 3D volume data, while limited cross-layer context further restricts representational capabilities and degrades performance.

[0007] Therefore, developing a low-dose, high-precision reconstruction method for medical images that is suitable for the characteristics of medical images has become an urgent need to improve the imaging quality of nuclear medicine images, reduce radiation exposure, and enhance the medical experience. Summary of the Invention

[0008] This invention provides a low-dose nuclear medicine PET image reconstruction method and system to overcome the deficiencies in the prior art.

[0009] In a first aspect, the present invention provides a method for reconstructing low-dose nuclear medicine PET images, comprising: Acquire low-dose nuclear medicine PET images and corresponding full-dose PET images to construct a training dataset; A low-dose PET reconstruction model based on a U-shaped encoder-decoder architecture was constructed. In the encoding stage, the input low-dose PET image features were decoupled in the frequency domain, decomposing them into low-frequency and high-frequency components. A global state-space model was used for global context modeling of the low-frequency components to suppress distribution degradation, while a local state-space model based on a space-filling curve scanning strategy was used for serialization modeling of the high-frequency components to maintain spatial locality and restore structural details. In the decoding stage, a full-dose semantic awareness cue branch was introduced. This involved aligning the feature distribution of the full-dose PET semantic features extracted from the pre-trained visual model with the low-dose PET semantic features, and then fusing the aligned semantic features with the decoder's multi-scale features to guide the recovery of low-dose features. The low-dose PET image is input into the reconstruction model, and the model is trained under supervision by a combination loss function to obtain a trained reconstruction model. The low-dose PET image to be reconstructed is input into the trained reconstruction model to obtain the reconstructed PET image.

[0010] According to the low-dose nuclear medicine PET image reconstruction method provided by the present invention, the frequency domain decoupling is achieved by discrete wavelet transform, which decomposes the input features in the encoding stage into multiple low-frequency sub-bands and multiple high-frequency sub-bands; the low-frequency sub-bands encode the overall structural information, and the high-frequency sub-bands encode the fine detail information.

[0011] According to the low-dose nuclear medicine PET image reconstruction method provided by the present invention, the global state space model adopts a bidirectional three-dimensional state space model, which simultaneously scans the flattened low-frequency feature sequence along both forward and backward dimensions, establishes long-range interactions between all spatial voxels, and realizes global context modeling of low-frequency degradation.

[0012] According to the low-dose nuclear medicine PET image reconstruction method provided by the present invention, the space-filling curve scanning strategy adopts a three-dimensional Hilbert curve. The three-dimensional spatial coordinates are converted into a one-dimensional index along a continuous trajectory that maintains local affinity through the Hilbert mapping function, so that adjacent elements in the one-dimensional sequence correspond to closely local regions in the spatial domain during the sequence modeling process.

[0013] According to the low-dose nuclear medicine PET image reconstruction method provided by the present invention, the full-dose semantic perception cue branch includes: extracting semantic features of low-dose PET images and full-dose PET images respectively using a frozen pre-trained visual model; The low-dose semantic features and the full-dose semantic features are aligned by cosine similarity constraint; the aligned semantic features are then fused with the multi-scale features of each stage of the decoder through a channel cross-attention mechanism.

[0014] According to the low-dose nuclear medicine PET image reconstruction method provided by the present invention, the feature distribution alignment further includes a domain adaptation strategy, which includes channel expansion and normalization of each slice of the low-dose PET image using statistics based on a natural image dataset to bridge the domain differences between natural images and medical images.

[0015] According to the present invention, a low-dose nuclear medicine PET image reconstruction method is provided, wherein the combined loss function includes a pixel-level loss for constraining the intensity information between the restored image and the full-dose image, a gradient loss for constraining the structural details between the restored image and the full-dose image, and a cosine similarity loss for constraining the alignment of low-dose semantic features to full-dose semantic features.

[0016] According to the present invention, a low-dose nuclear medicine PET image reconstruction method is provided, wherein the training dataset contains only paired data of low-dose PET images and full-dose PET images under a single dose reduction factor, and the reconstruction model after training is used for low-dose PET image reconstruction under various different dose reduction factors.

[0017] Secondly, the present invention also provides a low-dose nuclear medicine PET image reconstruction system, comprising: The acquisition module is used to acquire low-dose nuclear medicine PET images and corresponding full-dose PET images to construct a training dataset; A module is constructed to build a low-dose PET reconstruction model based on a U-shaped encoder-decoder architecture. In the encoding stage, the input low-dose PET image features are decoupled in the frequency domain, decomposing the features into low-frequency and high-frequency components. A global state-space model is used for global context modeling of the low-frequency components to suppress distribution degradation, while a local state-space model based on a space-filling curve scanning strategy is used for serialization modeling of the high-frequency components to maintain spatial locality and restore structural details. In the decoding stage, a full-dose semantic awareness cue branch is introduced. This involves aligning the feature distribution of the full-dose PET semantic features extracted from the pre-trained visual model with the low-dose PET semantic features, and then fusing the aligned semantic features with the decoder's multi-scale features to guide the recovery of low-dose features. The training module is used to input the low-dose PET image into the reconstruction model, and to supervise the training of the model by combining loss functions to obtain a trained reconstruction model. The output module is used to input the low-dose PET image to be reconstructed into the trained reconstruction model to obtain the reconstructed PET image.

[0018] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the low-dose nuclear medicine PET image reconstruction method as described above.

[0019] This invention provides a low-dose nuclear medicine PET image reconstruction method and system. It constructs a low-dose PET reconstruction model based on a typical U-shaped encoder-decoder structure, including multiple downsampling stages, a bottleneck stage, and an upsampling path symmetrical to the encoder. The downsampling stages combine frequency domain decoupling and a state-space model (SSM) to construct global and local models for high and low frequency subbands respectively. A turntable space model is used to model global context information in low-frequency components, while Hilbert fill curves are used to extract local relative relationships in high-frequency components. Furthermore, to align the dose information of the low-dose PET image with the full dose, this invention utilizes the powerful visual encoding capabilities of DINOV2 to design a feature modulation model combined with a cosine similarity loss function to align the feature distributions of low-dose PET with the full dose as a full-dose semantic perception cue branch. Before outputting the reconstructed PET image, a feature enhancement module is used to enhance the decoded low-dose PET features. Finally, a convolution output is used to recover the PET image and add it to the input low-dose PET for residual learning denoising. In addition, a comprehensive comparison of the processing results of this invention with existing state-of-the-art methods shows that the method of this invention achieves superior results. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a schematic flowchart of the low-dose nuclear medicine PET image reconstruction method provided by the present invention; Figure 2 This invention provides the overall framework for a low-dose nuclear medicine PET image intelligent reconstruction method based on a frequency domain decoupled visual state space model. Figure 3 This is a structural diagram of the frequency domain decoupled state space model provided by the present invention; Figure 4 This is a structural diagram of the full-dose semantic perception prompting and channel-level cross-attention mechanism provided by the present invention; Figure 5This is a schematic diagram of representative low-dose PET reconstruction results provided by the present invention; Figure 6 This is a schematic diagram of the low-dose nuclear medicine PET image reconstruction system provided by the present invention; Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] To address the problems of existing technologies, this invention presents an intelligent reconstruction method for low-dose nuclear medicine PET images based on a frequency-domain decoupled visual state-space model (SSM). This method designs a feature encoder based on the frequency-domain decoupled state-space model to balance low-frequency global contextual features with high-frequency local feature extraction. The global state-space module models low-frequency features, achieving effective recovery of low-frequency degradation information. For high-frequency components, leveraging the characteristic of Hilbert curves to maintain the continuity of local spatial neighborhoods, a local state-space module based on a Hilbert scanning strategy is designed to sequentially model high-frequency local features, improving the model's ability to recover fine-grained structural details while preserving local structural consistency. The full-dose perception cue branch utilizes the powerful prior information perception capability of the visual pre-trained model DINOv2 combined with a channel-level cross-attention mechanism. Through a feature modulation model combined with cosine similarity, it aligns the feature distributions of low-dose PET with those of standard dose PET, thereby improving the generalization ability of the entire model. Extensive experiments on low-dose PET medical image datasets under various public and private DRFs demonstrate that the method of this invention achieves significant improvements in structural fidelity and quantitative accuracy compared to several baseline methods.

[0024] Figure 1 This is a schematic flowchart of the low-dose nuclear medicine PET image reconstruction method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, it includes: Step 100: Acquire low-dose nuclear medicine PET images and corresponding full-dose PET images to construct a training dataset; Step 200: Construct a low-dose PET reconstruction model based on a U-shaped encoder-decoder architecture. In the encoding stage, the input low-dose PET image features are decoupled in the frequency domain, decomposing the features into low-frequency and high-frequency components. A global state-space model is used for global context modeling of the low-frequency components to suppress distribution degradation, while a local state-space model based on a space-filling curve scanning strategy is used for serialization modeling of the high-frequency components to maintain spatial locality and restore structural details. In the decoding stage, a full-dose semantic awareness cue branch is introduced. The full-dose PET semantic features extracted from the pre-trained visual model are aligned with the low-dose PET semantic features, and the aligned semantic features are fused with the multi-scale features of the decoder to guide the recovery of low-dose features. Step 300: Input the low-dose PET image into the reconstruction model, and perform supervised training on the model by combining loss functions to obtain a trained reconstruction model; Step 400: Input the low-dose PET image to be reconstructed into the trained reconstruction model to obtain the reconstructed PET image.

[0025] Specifically, embodiments of the present invention include: (1) A private dataset was obtained by scanning with a GE PET scanner (Discovery 710) and 68Ga-FAPI under normal scanning conditions. All data were acquired in list mode, allowing for data rebinding to simulate different acquisition times (1 / 2, 1 / 4, 1 / 10, and 1 / 20 of the original dose). All images were converted to Standardized Uptake Value (SUV). Low-dose PET images were reconstructed using the Ordered Subset Expectation-Maximization (OSEM) algorithm with 2 iterations, 24 subsets, and a point spread function (PSF) of 6.4 mm to generate reference images. The matrix and voxel sizes were set to [values ​​to be filled in]. and mm³. Each simulated low-dose PET image was reconstructed by resampling a specific time window during acquisition and combining it with corresponding time reduction to obtain a low-dose PET image with a certain dose reduction factor (DRF). The public dataset was acquired using a Siemens Biograph VisionQuadra PET / CT scanner. Image reconstruction employed the Ordered Subset Expectation-Maximization (OSEM) algorithm (4 iterations, 5 subsets), combined with Time-of-Flight (TOF) and Point Spread Function (PSF) modeling. After reconstruction, a Gaussian filter with a full width at half maximum (FWHM) of 2 mm was applied to obtain an isotropic voxel size of 1.65 mm³. The original data was recorded in list mode to facilitate simulation of different dose levels through time resegmentation bins. Specifically, the low-dose PET images were generated by subsampling the original full-dose list mode data using different dose reduction factors (DRFs) of 4, 10, 20, 50, and 100. All generated images were converted to normalized uptake values ​​(SUV) units. All low-dose PET and standard-dose PET images were divided into training, validation, and test datasets in a 7:1:2 ratio.

[0026] (2) A low-dose PET reconstruction model based on a typical U-shaped encoder-decoder structure is constructed, including multiple downsampling stages, a bottleneck stage, and an upsampling path symmetrical to the encoder. Furthermore, to align the dose information of the low-dose PET image with the full dose, this invention utilizes the powerful visual encoding capabilities of DINOV2 to design a feature modulation model combined with a cosine similarity loss function to align the feature distributions of low-dose PET with the full dose as a full-dose semantic perception cue branch. Before outputting the reconstructed PET image, a feature enhancement module is used to enhance the decoded low-dose PET features. Finally, a... The convolutional output recovers the PET image and adds it to the input low-dose PET image for residual learning denoising.

[0027] (3) Low-dose PET images under different DRF conditions are input into the aforementioned low-dose reconstruction model, and the multimodal medical image fusion framework is supervised through various loss functions. Specifically, this invention employs... Both gradient loss and gradient loss are used to ensure that the intensity information and structural details of the recovered PET are consistent. Furthermore, a cosine similarity loss is introduced in the full-dose semantic awareness cue branch to constrain the alignment of low-dose semantic information with full-dose features.

[0028] (4) Input the low-dose PET images under the above DRFs into the trained low-dose reconstruction model to obtain the reconstructed PET images.

[0029] Furthermore, the State-Space Model (SSM) and Hilbert curve are as follows: State-space models (SSMs): At their core is a linear time-invariant system. Used for sequence-to-sequence modeling, they possess dynamic properties that remain constant over time. Due to their linear complexity, SSMs can model one-dimensional functions or sequences (…). Implicit mapping to the latent state space ( ), and then mapped to the output ( This allows for the effective capture of the system's intrinsic dynamics. Formally, the SSM can be expressed as the following linear ordinary differential equation (ODE):

[0030] in, This represents the state matrix. and These represent the mapping parameters.

[0031] To integrate continuous-time SSMs into deep learning frameworks for vision tasks, the underlying ordinary differential equations must be discretized. Specifically, this is achieved by introducing a learnable time-scale parameter. Using the Zero-Order Hold (ZOH) rule, continuous parameters are... and Convert them to their discrete corresponding values, denoted as and The conversion formula is as follows:

[0032] After discretization, discrete SSM can be evaluated using two equivalent computational models: linear recursion and global convolution.

[0033] Linear recursion:

[0034] Global convolution:

[0035] in Let L represent an input sequence of length L. Represents a structured convolution kernel. This represents the convolution operation.

[0036] Unlike traditional Sequence-State Models (SSMs) that rely on fixed transition parameters (e.g., A, B), the S6 mechanism uses B, C, and... Input-dependent parameterization is introduced. This dynamic formulation enables the model to continuously adapt to the sequence context and effectively capture content-aware representations.

[0037] Hilbert curves, as a representative space-filling curve (SFC), are renowned for preserving strong spatial locality when mapping multidimensional data to one-dimensional sequences. Given their excellent clustering performance in three-dimensional space, this invention employs Hilbert scanning to process three-dimensional low-dose PET images. This strategy significantly enhances inter-slice correlation by preserving anatomical locality in sequence representation. Mathematically, d-dimensional SFC is derived from the mapping... Definition. For discrete 3D PET volume data, the curve order n is related to the spatial resolution. Therefore, for any two sequence indices... The spatial linear ratio (SLR) can be expressed as follows:

[0038] The dilation factor (DF) is defined as the supremum of SLR and is used to evaluate the ability of SFC to preserve spatial locality. A significantly lower DF value ensures that neighboring elements in a one-dimensional sequence map to tightly localized regions in the spatial domain. Theoretically, the DF value of the Hilbert curve is constant at 6, while the DF value of conventional raster scanning tends to infinity as the curve order n increases. This mathematical divergence ingeniously inspires the frequency decoupling architecture proposed in this invention: the global dispersion of raster scanning is suitable for capturing long-range dependencies, while the bounded DF value of Hilbert is well-suited for strict locality preservation. By specifically applying 3D Hilbert scanning to the high-frequency components of low-dose PET, the low-dose reconstruction model effectively preserves the fine local texture in the spatial structure model (SSM), thus providing a powerful complement to global modeling applied to low-frequency features.

[0039] Furthermore, the plurality of downsampling stages include: Input embedding layer: consisting of a Convolutional layers map low-dose PET input images to a high-dimensional feature space.

[0040] Frequency Decoupling SSM Module: This invention employs Discrete Wavelet Transform (DWT) to decompose low-dose and full-dose PET images into low-frequency and high-frequency components, followed by cross-spectral reconstruction. Frequency domain analysis shows that noise degradation in low-dose PET images is mainly concentrated in the low-frequency band, while the high-frequency band primarily encodes structural details and exhibits relatively less degradation. Based on these observations, this invention introduces a frequency decoupling modeling strategy in the frequency decoupling SSM module. Specifically, a global three-dimensional spatial structure model (SSM) is deployed for low-frequency features to capture long-range dependencies and effectively suppress global distribution degradation. Conversely, to handle high-frequency components, a local SSM utilizing the Hilbert scan strategy is designed. This strategy strictly maintains spatial locality during sequence modeling, thereby contributing to the accurate reconstruction of fine structural details. Formally, the formula for this frequency decoupling modeling strategy is as follows:

[0041] in, They represent the first class( The input and output characteristics of the frequency decoupling module. After DWT decomposition, the characteristics are decomposed into low-frequency components. and high frequency components , here These represent the height, length, and width of the input low-dose PET image, respectively. Specifically, in sub-band representation (e.g., ,…, )middle, and These represent low and high frequencies, respectively. Furthermore, 3 and These represent the global three-dimensional state space module and the local state space module guided by the Hilbert scan strategy, respectively.

[0042] 3 It is a bidirectional state-space model. To handle low-frequency features, this invention performs a global flattening operation along both the forward (DHW) and backward (WHD) dimensions. Input features It is then routed to a standard SSM module, which has a bidirectional... 3 Layer. By simultaneously scanning the linearized voxel sequence from both directions, this bidirectional mechanism explicitly establishes long-range interactions among all spatial voxels, thereby enabling robust global context modeling for low-frequency degradation. The output sequence is then reconstructed back to its original spatial dimension to obtain globally context-dependent low-frequency features. .

[0043] The HBSSM module is designed to utilize the locality-preserving properties of 3D Hilbert scans to process high-frequency components. Specifically, each spatial coordinate Through Hilbert mapping function This is converted to a one-dimensional index, thus flattening the 3D space along a continuous trajectory that preserves local affinity. Following this path, the input high-frequency subband features... By downsampling to a fixed spatial resolution using adaptive average pooling, we obtain... The pooled features are then flattened into a locally enhanced one-dimensional sequence. The enhanced sequence is processed by the same SSM module as the global module. This step explicitly captures fine-grained high-frequency textures at the local level. Finally, the enhanced sequence is reshaped and interpolated back to the original input resolution to construct locally refined high-frequency features. .

[0044] Furthermore, to bridge the distribution gap between low-dose and full-dose PET images, the full-dose semantic-aware cue branch introduces a semantically adaptive modulation module driven by a pre-trained DINOv2 model. This module is equipped with cosine similarity constraints to explicitly align feature distributions across different dose levels. Subsequently, a channel cross-attention mechanism is deployed in the decoder to fuse these modulated prior cues with multi-scale features from the encoder module. This architecture significantly enhances the model's generalization ability, which is particularly important in extremely low-dose scenarios.

[0045] Semantic Adaptive Modulation Module: Enhancing DINOv2 Features with the Adapter and decoder features As the primary input, low-dose PET slices are processed by a frozen DINOv2 model and subsequently reconstructed into a 3D volumetric representation in the full-dose semantic-aware cueing branch. To address the inherent domain differences between natural images and PET data, this invention implements a domain-adaptive strategy that includes channel expansion followed by normalization of low-dose PET slices per layer using ImageNet-based statistics. This alignment ensures that semantic priors can be effectively transferred to the medical domain. Subsequently, DINOv2 features are refined using an adapter built on the SwiGLU architecture. The corresponding full-dose features... The same process is used, but excluding the SwiGLU adapter. Finally, to enable full-dose semantic awareness in the low-dose representation, this invention uses cosine similarity constraints for explicit alignment. and The distribution of is given by the following formula:

[0046] in, and These represent the volumes of low-dose and full-dose PET, respectively. This represents the cosine similarity loss. and Representing the spatial elements of the index feature sequence and respectively Normalization. Subsequently... The process proceeds after a Layer Normalization (LN) layer and two stacked combination layers (each combination layer containing a convolutional layer, an instance normalization (IN) layer, and a LeakyReLU (LReLU) activation layer). The first convolution uses a stride of... The first convolution kernel, the second convolution uses The convolution kernel is used to obtain intermediate features. To make the upsampled skip connection features The spatial resolution is matched with the subsequent feature modulation. Alignment is achieved through trilinear interpolation. The generation process is as follows:

[0047] In this formula, , ,as well as These represent the convolution kernel and stride, respectively. The convolution operation, LeakyReLU activation function, and trilinear interpolation are then performed. After parallel Convolutional layers yield modulation parameters and By using the Sigmoid activation function... The input features are explicitly constrained within the interval [0,1]. The final semantic representation is obtained by modulating the input features. .

[0048] Furthermore, the channel cross-attention mechanism includes: To inject full-dose semantic awareness into the decoding stage and effectively guide low-dose feature recovery while suppressing noise artifacts, this invention introduces a Channel-wise Cross-Attention (CCA) module in each decoding stage. The main decoder features are processed through 1×1×1 and 3×3×3 depthwise convolutions, then reshaped to generate the query. ,in At the same time, in order to eliminate domain offset, Preprocessing is performed using LN layers, followed by mapping using corresponding convolutional layers to generate keys. Sum The mathematical expression for the CCA module is as follows:

[0049] in, It is a learnable scaling factor, applied to and The dot product between them enhances the flexibility of attention distribution before the Softmax operation.

[0050] Furthermore, the model training process employs the following optimization strategy: Training input resolution: The input 3D PET image is randomly cropped into sizes of [sizes to be filled in]. and Sub-images.

[0051] Learning rate scheduling: The AdamW optimizer is used, the initial learning rate is set to 0.0001, and there is no learning rate optimization scheduling; Gradient clipping: The global gradient norm is limited to 1.0 to prevent training instability.

[0052] Mixed Precision Training and Data Augmentation: Automatic Mixed Precision (AMP) is employed to accelerate computation and reduce memory overhead. Furthermore, random spatial flipping is performed in real-time along the depth, height, and width axes to effectively mitigate overfitting.

[0053] Furthermore, the training process of the low-dose nuclear medicine PET image reconstruction model includes: acquiring normal-dose nuclear medicine PET volumetric images under conventional scanning conditions and low-dose PET images (including private and public data) reconstructed under different DRFs using list mode. These are then divided into training, validation, and test datasets in a 7:1:2 ratio.

[0054] The technical solution of the present invention will be illustrated below with a specific embodiment.

[0055] S1: Volumetric PET images of normal dose acquired under standard scanning conditions, and low-dose PET images reconstructed using different DRFs in list mode (including private and public data). Low-dose and full-dose image pairs were divided into training, validation, and testing pairs in a 7:1:2 ratio. S2: A low-dose nuclear medicine PET image reconstruction method based on a frequency domain decoupled visual state-space model (SSM) is constructed. This model includes a feature downsampling encoding stage, a bottleneck stage, and a feature upsampling decoding stage. The feature downsampling encoding stage encodes features from the input nuclear medicine PET images under different DRFs, outputting feature maps for four stages. To align the dose information of the low-dose PET images with the full dose, this invention utilizes the powerful visual encoding capabilities of DINOV2 to design a feature modulation model combined with a cosine similarity loss function to align the feature distributions of low-dose PET with the full dose, serving as a full-dose semantic perception cue branch. Then, in the feature decoding stage, the features from the full-dose semantic perception cue branch interact with the upsampled features to progressively recover the input feature size. Before outputting the reconstructed PET image, a feature enhancement module is used to enhance the decoded low-dose PET features. Finally, a... The convolutional output recovers the PET image and adds it to the input low-dose PET image for residual learning denoising; S3: Low-dose PET images under different DRF conditions are input into the aforementioned low-dose reconstruction model, and the multimodal medical image fusion framework is supervised through various loss functions. Specifically, this invention employs loss and gradient loss to ensure the intensity information and structural details of the recovered PET images are consistent. Furthermore, cosine similarity loss is introduced in the full-dose semantic awareness cue branch to constrain the alignment of low-dose semantic information with full-dose features; S4: Input the low-dose PET images under various DRF conditions described above into the trained low-dose reconstruction model to obtain the reconstructed PET images. The objective performance of the low-dose PET model is evaluated using three metrics commonly used for image restoration assessment: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Normalized Root Mean Square Error (NRMSE).

[0056] In the specific implementation process, in S1, clinical PET data of 79 patients were collected from a molecular imaging PET-CT / MRI center of a certain hospital. This cohort included 44 women and 35 men, with a mean age of 62 ± 13 years. Specifically, a GE Discovery 710 PET scanner and 68 Ga-FAPI radiotracers were used for target scanning. The original list pattern data was accurately re-binded to simulate shortened acquisition times corresponding to DRFs of 2, 4, 10, and 20 (i.e., 1 / 2, 1 / 4, 1 / 10, and 1 / 20 of the full dose, respectively). A reference full-dose image was reconstructed using the OSEM algorithm, configured with 2 iterations, 24 subsets, and a 6.4 mm PSF filter. matrix size and The volume was in mm³ voxel size. Each simulated low-dose PET volume was reconstructed from counts extracted from the acquisition center time window and scaled proportionally to the shortened duration. All volumes were normalized to SUV units. Subsequently, they were allocated in a 7:1:2 ratio to 55 pairs (DRF=20) for training, 8 pairs for validation, and 16 pairs for testing.

[0057] To validate the model's performance across different modalities, a dataset from the Ultra-Low Dose PET Imaging Challenge was selected. This dataset contains images of 377 subjects acquired using a Siemens Biograph Vision Quadra PET / CT scanner. 18 F-FDG PET scan images. Image reconstruction was performed using the ordered subset expectation-maximization (OSEM) algorithm (4 iterations, 5 subsets), combined with time-of-flight (TOF) and point spread function (PSF) modeling. After reconstruction, a Gaussian filter with a full width at half maximum (FWHM) of 2 mm was applied [1] to obtain 1.65 mm. 3The isotropic voxel sizes were determined. The raw data was recorded in list mode to facilitate simulation of different dose levels using time-division binning. Specifically, low-dose PET images were generated by subsampling the raw full-dose list mode data using five different DRFs: 4, 10, 20, 50, and 100. All generated images were converted to normalized uptake values ​​(SUVs). The dataset was divided into 297 pairs of full-dose and low-dose (DRF=20) images for training, 20 pairs for validation, and the remaining 60 pairs for testing. Furthermore, to verify the model's robustness under different noise distributions, the test set also included images reconstructed under the other four DRFs. These test samples were selected to highlight different features and evaluate the reconstruction performance of the method under different tracers.

[0058] In step S2, the frequency domain decoupled visual state-space model (SSM) includes a feature downsampling encoding stage, a bottleneck stage, and a feature upsampling decoding stage. Specifically, the feature downsampling encoding stage encodes features from nuclear medicine PET images under different DRFs, outputting feature maps for four stages. Each feature is obtained by constructing global and local models for high and low frequency subbands using the frequency domain decoupled state-space model (SSM). A turntable space model is used to model global context information in the low-frequency components, while Hilbert fill curves are used to extract local relative relationships in the high-frequency components. To align the dose information of low-dose PET images with the full dose, this invention utilizes the powerful visual encoding capabilities of DINOV2 to design a feature modulation model combined with a cosine similarity loss function to align the feature distributions of low-dose PET with the full dose, serving as a full-dose semantic perception cue branch.

[0059] like Figure 2 The overall framework of the intelligent reconstruction method for low-dose nuclear medicine PET images based on the frequency domain decoupled visual state-space model (SSM) is implemented in steps. First, low-dose PET images under different DRFs are used to extract features through a multi-stage encoder. Then, the low-dose PET images under the input DRF are fed into a full-dose semantic awareness cue molecule at the 2D slice level, combined with a cosine similarity loss function to align the feature distributions of low-dose PET images with those of the full dose. In the feature decoding stage, the input feature size is gradually restored by interacting with the features of the full-dose semantic awareness cue branch and the upsampled features. Before outputting the reconstructed PET image, a feature enhancement module is used to enhance the features of the decoded low-dose PET images. Finally, the reconstructed PET image is output using convolutional layers and residual learning.

[0060] like Figure 3The frequency-domain decoupled visual state-space model (SSM) is described: Discrete Wavelet Transform (DWT) is used to decompose low-dose and full-dose PET images into low-frequency and high-frequency components, followed by cross-spectral reconstruction. Frequency domain analysis shows that noise degradation in low-dose PET images is mainly concentrated in the low-frequency band, while the high-frequency band mainly encodes structural details and exhibits relatively less degradation. Based on these observations, this invention introduces a frequency decoupling modeling strategy in the frequency-decoupled SSM module. Specifically, a global three-dimensional spatial structure model (SSM) is deployed for low-frequency features to capture long-range dependencies and effectively suppress global distribution degradation. Conversely, to handle high-frequency components, a local SSM utilizing the Hilbert scan strategy is designed. This strategy strictly maintains spatial locality during sequence modeling, thereby contributing to the accurate reconstruction of fine structural details.

[0061] like Figure 4 The full-dose semantic awareness cue branch, as illustrated by the adaptive semantic awareness module and channel-level cross-attention mechanism, aims to bridge the distribution gap between low-dose and full-dose PET images. It introduces a semantically adaptive modulation module driven by a pre-trained DINOv2 model. This module is equipped with cosine similarity constraints to explicitly align feature distributions across different dose levels. Subsequently, a channel-level cross-attention mechanism is deployed in the decoder to fuse these modulated prior cues with multi-scale features from the encoder module. This architecture significantly enhances the model's generalization ability.

[0062] Clinical data collected at a PET-CT / MRI center of the Molecular Imaging Center at Renmin Hospital of Wuhan University under different diffusion resonant frequencies (DRFs) were selected for the test set. PET images under each DRF included full-dose and low-dose PET reconstruction results, including 3D UNet (a CNN-based method), 3D GAN (a generative adversarial network-based method), Restormer, DRMC, SpachTransforemr (a Transforemr-based method), CoreDiff (a 2.5D diffusion model method), MAIR and TAT (2D methods), and the method of this invention. The reference metrics were Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Normalized Root Mean Square Error (NRMSE). These metrics evaluate the quality of PET reconstruction from different perspectives. PSNR assesses voxel-level reconstruction fidelity, while SSIM simultaneously assesses structural similarity and perceived image quality. NRMSE is used to assess voxel-level intensity accuracy. Tables 1 and 2 present the objective comparison results of various methods. The method of this invention has superior reconstruction performance and the highest degree of preservation of detailed structure in low-dose PET images. In particular, the PSNR, SSIM, and NRMSE indices show that the method of this invention has significant advantages in terms of reconstruction fidelity, structural similarity to full-dose PET, and voxel-level intensity accuracy. Figure 5 The fusion results of the method of this invention with other baseline methods under different DRFs are shown. To more clearly demonstrate image details, regions of interest (ROIs) or arrows are selected to show the reconstruction quality of key regions. "Ours" represents the reconstruction result of this invention, and "unseen" indicates testing only under this DRF. Compared to other comparison methods, the reconstruction results obtained by the method of this invention can better restore image details, and the subjective results are more reliable.

[0063] Table 1. Objective index results of various comparison methods under DRF20, with the best results shown in bold.

[0064] Table 2 shows the objective index results of various comparison methods under DRF10, with the best results displayed in bold.

[0065] The low-dose nuclear medicine PET image reconstruction system provided by the present invention is described below. The low-dose nuclear medicine PET image reconstruction system described below can be referred to in correspondence with the low-dose nuclear medicine PET image reconstruction method described above.

[0066] Figure 6 This is a schematic diagram of the structure of the low-dose nuclear medicine PET image reconstruction system provided in an embodiment of the present invention, as shown below. Figure 6As shown, it includes: acquisition module 61, construction module 62, training module 63, and output module 64, wherein: The acquisition module 61 is used to acquire low-dose nuclear medicine PET images and corresponding full-dose PET images to construct a training dataset. The construction module 62 is used to construct a low-dose PET reconstruction model based on a U-shaped encoder-decoder architecture. In the encoding stage, the input low-dose PET image features are decoupled in the frequency domain and decomposed into low-frequency and high-frequency components. The low-frequency components are modeled using a global state-space model to suppress distribution degradation, and the high-frequency components are modeled using a local state-space model based on a space-filling curve scanning strategy to maintain spatial locality and restore structural details. In the decoding stage, a full-dose semantic perception cue branch is introduced. The full-dose PET semantic features extracted by the pre-trained visual model are aligned with the low-dose PET semantic features, and the aligned semantic features are fused with the multi-scale features of the decoder to guide the recovery of low-dose features. The training module 63 is used to input the low-dose PET images into the reconstruction model and supervise the model training by combining loss functions to obtain a trained reconstruction model. The output module 64 is used to input the low-dose PET images to be reconstructed into the trained reconstruction model to obtain the reconstructed PET images.

[0067] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a low-dose nuclear medicine PET image reconstruction method. This method includes: acquiring low-dose nuclear medicine PET images and corresponding full-dose PET images, constructing a training dataset; building a low-dose PET reconstruction model based on a U-shaped encoder-decoder architecture; wherein, in the encoding stage, the input low-dose PET image features are decoupled in the frequency domain, decomposing the features into low-frequency and high-frequency components; a global state-space model is used for global context modeling of the low-frequency components to suppress distribution degradation; a local state-space model based on a space-filling curve scanning strategy is used for serialization modeling of the high-frequency components to maintain spatial locality and restore structural details; in the decoding stage, a full-dose semantic perception cue branch is introduced, aligning the feature distribution of the full-dose PET semantic features extracted by the pre-trained visual model with the low-dose PET semantic features, and fusing the aligned semantic features with the multi-scale features of the decoder to guide the recovery of low-dose features; the low-dose PET image is input into the reconstruction model, and the model is trained under supervision using a combined loss function to obtain a trained reconstruction model; the low-dose PET image to be reconstructed is input into the trained reconstruction model to obtain the reconstructed PET image.

[0068] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, 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 described in the various embodiments of the present invention. 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.

[0069] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0070] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reconstructing low-dose nuclear medicine PET images, characterized in that, include: Acquire low-dose nuclear medicine PET images and corresponding full-dose PET images to construct a training dataset; A low-dose PET reconstruction model based on a U-shaped encoder-decoder architecture was constructed. In the encoding stage, the input low-dose PET image features were decoupled in the frequency domain, decomposing them into low-frequency and high-frequency components. A global state-space model was used for global context modeling of the low-frequency components to suppress distribution degradation, while a local state-space model based on a space-filling curve scanning strategy was used for serialization modeling of the high-frequency components to maintain spatial locality and restore structural details. In the decoding stage, a full-dose semantic awareness cue branch was introduced. This involved aligning the feature distribution of the full-dose PET semantic features extracted from the pre-trained visual model with the low-dose PET semantic features, and then fusing the aligned semantic features with the decoder's multi-scale features to guide the recovery of low-dose features. The low-dose PET image is input into the reconstruction model, and the model is trained under supervision by a combination loss function to obtain a trained reconstruction model. The low-dose PET image to be reconstructed is input into the trained reconstruction model to obtain the reconstructed PET image.

2. The low-dose nuclear medicine PET image reconstruction method according to claim 1, characterized in that, The frequency domain decoupling is achieved through discrete wavelet transform, which decomposes the input features in the coding stage into multiple low-frequency sub-bands and multiple high-frequency sub-bands; the low-frequency sub-bands encode the overall structural information, and the high-frequency sub-bands encode the fine detail information.

3. The low-dose nuclear medicine PET image reconstruction method according to claim 1, characterized in that, The global state space model adopts a bidirectional three-dimensional state space model, which simultaneously scans the flattened low-frequency feature sequence along both forward and backward dimensions, establishes long-range interactions between all spatial voxels, and realizes global context modeling of low-frequency degradation.

4. The low-dose nuclear medicine PET image reconstruction method according to claim 1, characterized in that, The space-filling curve scanning strategy uses a three-dimensional Hilbert curve. The three-dimensional spatial coordinates are converted into a one-dimensional index along a continuous trajectory that maintains local affinity through the Hilbert mapping function, so that adjacent elements in the one-dimensional sequence correspond to closely local regions in the spatial domain during sequence modeling.

5. The low-dose nuclear medicine PET image reconstruction method according to claim 1, characterized in that, The full-dose semantic perception prompting branch includes: extracting semantic features from low-dose PET images and full-dose PET images respectively using a frozen pre-trained visual model; The low-dose semantic features and the full-dose semantic features are aligned by cosine similarity constraint; the aligned semantic features are then fused with the multi-scale features of each stage of the decoder through a channel cross-attention mechanism.

6. The low-dose nuclear medicine PET image reconstruction method according to claim 5, characterized in that, The feature distribution alignment also includes a domain adaptation strategy, which includes channel expansion and normalization of slices of low-dose PET images using statistics based on natural image datasets to bridge the domain differences between natural and medical images.

7. The low-dose nuclear medicine PET image reconstruction method according to claim 1, characterized in that, The combined loss function includes a pixel-level loss for constraining intensity information between the restored image and the full-dose image, a gradient loss for constraining structural details between the restored image and the full-dose image, and a cosine similarity loss for constraining the alignment of low-dose semantic features to full-dose semantic features.

8. The low-dose nuclear medicine PET image reconstruction method according to any one of claims 1 to 7, characterized in that, The training dataset contains only paired data of low-dose PET images and full-dose PET images under a single dose reduction factor. The reconstruction model after training is used for low-dose PET image reconstruction under various dose reduction factors.

9. A low-dose nuclear medicine PET image reconstruction system, characterized in that, include: The acquisition module is used to acquire low-dose nuclear medicine PET images and corresponding full-dose PET images to construct a training dataset; A module is constructed to build a low-dose PET reconstruction model based on a U-shaped encoder-decoder architecture. In the encoding stage, the input low-dose PET image features are decoupled in the frequency domain, decomposing the features into low-frequency and high-frequency components. A global state-space model is used for global context modeling of the low-frequency components to suppress distribution degradation, while a local state-space model based on a space-filling curve scanning strategy is used for serialization modeling of the high-frequency components to maintain spatial locality and restore structural details. In the decoding stage, a full-dose semantic awareness cue branch is introduced. This involves aligning the feature distribution of the full-dose PET semantic features extracted from the pre-trained visual model with the low-dose PET semantic features, and then fusing the aligned semantic features with the decoder's multi-scale features to guide the recovery of low-dose features. The training module is used to input the low-dose PET image into the reconstruction model, and to supervise the training of the model by combining loss functions to obtain a trained reconstruction model. The output module is used to input the low-dose PET image to be reconstructed into the trained reconstruction model to obtain the reconstructed PET image.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the low-dose nuclear medicine PET image reconstruction method as described in any one of claims 1 to 8.