A method, apparatus and equipment for reconstructing PET images

Abstract

This specification relates to a method, apparatus, and device for reconstructing PET images. The method includes: acquiring a positron emission tomography (PET) image to be reconstructed; inputting the PET image into a pre-trained degradation pattern recognition model to obtain degradation pattern information of the PET image; inputting the PET image and the degradation pattern information into a pre-trained reconstruction model to obtain a reconstructed PET image; wherein the reconstruction model uses the degradation pattern information as prior information to predict and remove degradation components of the PET image. By identifying and utilizing degradation pattern information, the reconstruction model can achieve reconstruction quality no less than that of multimodal fusion methods using only single-modal PET data. This fundamentally avoids registration errors and significantly reduces patient radiation exposure and examination costs.

Description

Technical Field

[0001] This specification relates to the field of image processing technology, and in particular to a method, apparatus and device for reconstructing PET images. Background Technology

[0002] Positron emission tomography (PET) is an important functional molecular imaging technique that plays an irreplaceable role in the early diagnosis and treatment evaluation of tumors, neurological, and cardiovascular diseases. However, the radioactive tracers used in PET imaging inflict radiation doses on patients. To adhere to the principle of "the lowest possible appropriate dose," clinical practice urgently requires minimizing the injection dose while maintaining diagnostic image quality. However, reducing the dose directly leads to a decrease in photon count, causing severe image degradation problems, primarily manifested as a sharp drop in signal-to-noise ratio, significant noise and artifacts, severely affecting the quantitative accuracy of the images and the physician's clinical interpretation.

[0003] To reconstruct high-quality images from low-dose PET data, existing techniques rely on multimodal data for reconstruction, such as combining prior anatomical information provided by computed tomography (CT) or magnetic resonance imaging (MRI) to improve reconstruction quality.

[0004] However, such methods have the following limitations: Firstly, they lead to increased radiation dose and cost. CT modality introduces additional X-ray radiation, while MRI prolongs scan time and increases equipment costs. Secondly, registration complexity is high. Spatial registration errors between multimodal images can easily cause information misalignment, affecting reconstruction accuracy. Summary of the Invention

[0005] To address the problems in the prior art, embodiments of this specification provide a method, apparatus, and device for reconstructing PET images. By identifying the degradation mode information of the PET image to be reconstructed, the degradation mode information and the PET image to be reconstructed are simultaneously input into the reconstruction model, providing targeted prior guidance for the reconstruction model. This fully exploits the implicit global and local features in the PET image, achieving reconstruction quality no less than that of multimodal fusion methods using only single-modal PET data. This fundamentally avoids registration errors and significantly reduces patient radiation exposure and examination costs.

[0006] On one hand, embodiments of this specification provide a method for reconstructing PET images, the method comprising:

[0007] Acquire positron emission tomography (PET) images to be reconstructed;

[0008] The PET image is input into a pre-trained degradation pattern recognition model to obtain degradation pattern information of the PET image;

[0009] The PET image and the degradation pattern information are input into a pre-trained reconstruction model to obtain a reconstructed PET image; wherein the reconstruction model uses the degradation pattern information as prior information to predict and remove the degradation components of the PET image.

[0010] Furthermore, the degradation mode information includes physical parameter information related to the imaging process and state information related to image post-processing.

[0011] Furthermore, the degradation mode information includes tracer type, tracer dosage level, and correction status; the correction status indicates whether the PET image has been corrected using other modal data.

[0012] Furthermore, inputting the PET image and the degradation pattern information into the pre-trained reconstruction model further includes: encoding the degradation pattern information into a semantic feature vector; and inputting the semantic feature vector and the PET image into the reconstruction model.

[0013] The reconstruction model adjusts the spatial attention weights and / or channel attention weights of its attention mechanism based on the semantic feature vector.

[0014] Furthermore, the reconstruction model uses the degradation pattern information as prior information and predicts different types of degradation components through its multiple degradation component prediction branches.

[0015] Furthermore, the plurality of degradation component prediction branches include a branch for predicting Poisson noise components, a branch for predicting Gaussian noise components, and a branch for predicting physical artifact components.

[0016] Furthermore, the training sample data for the reconstruction model includes: standard PET images and noisy PET images;

[0017] The noisy PET image is obtained by the following steps: Based on the preset total time steps T, on the basis of the standard PET image, multiple degradation components corresponding to the physical mechanism of PET imaging are injected in stages through T consecutive noise addition steps to generate a continuous image sequence from a slightly noisy image to a completely noisy image.

[0018] Furthermore, the method also includes:

[0019] In each noise addition step, the following three types of degradation components are injected: Poisson noise component, Gaussian noise component, and physical artifact component.

[0020] Furthermore, the method also includes:

[0021] For each noise addition step, the true value of each type of degradation component injected is recorded during the process from the standard PET image to the current noisy PET image, and used as a supervision label for training the reconstruction model.

[0022] On the other hand, this specification also provides a method for training a reconstruction model for reconstructing PET images, the method comprising:

[0023] Obtain a set of training sample pairs, wherein each training sample pair includes a standard PET image and a noisy image corresponding to a different time step, generated by performing a noisy step on the standard PET image;

[0024] Based on the training sample pair set, the initialized reconstruction model is iteratively optimized;

[0025] In each iteration, the following steps are performed:

[0026] A training sample pair is sampled from the training sample pair set, the sample pair containing a standard PET image and a noisy PET image corresponding to the first time step;

[0027] The noisy PET image corresponding to the first time step and the first time step are input into the reconstruction model to obtain the predicted values ​​of the reconstruction model for multiple degradation components in the noisy PET image.

[0028] The loss function is calculated based on the difference between the predicted value and the actual degradation component value corresponding to the first time step;

[0029] The parameters of the reconstruction model are updated based on the loss function.

[0030] On the other hand, embodiments of this specification provide a PET image reconstruction apparatus, the apparatus comprising:

[0031] The acquisition module is used to acquire the positron emission tomography (PET) image to be reconstructed.

[0032] The recognition module is used to input the PET image into a pre-trained degradation pattern recognition model to obtain degradation pattern information of the PET image;

[0033] The reconstruction module is used to input the PET image and the degradation pattern information into a pre-trained reconstruction model to obtain a reconstructed PET image; wherein the reconstruction model uses the degradation pattern information as prior information to predict and remove the degradation components of the PET image.

[0034] On the other hand, embodiments of this specification also provide a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the above-described method.

[0035] The advantages of the methods described in this specification include, but are not limited to:

[0036] (1) By identifying the degradation mode information of the PET image to be reconstructed, the degradation mode information and the PET image to be reconstructed are simultaneously input into the reconstruction model, providing targeted prior guidance for the reconstruction model, thereby fully exploring the hidden global and local features in the PET image, significantly improving the signal-to-noise ratio and diagnostic reliability of low-dose PET images, and finally obtaining high-quality and high-stability reconstruction results.

[0037] (2) It solves the technical problem that mainstream models have poor dynamic adaptability to tracer types and dosage levels, and need to be trained independently for different scenarios. It significantly improves the generalization of reconstruction models, enhances dynamic adaptability, and reduces model training costs.

[0038] (3) Existing reconstruction methods heavily rely on CT or MRI images to provide prior anatomical information to aid reconstruction. This not only introduces additional radiation doses or scan time and costs, but may also lead to reconstruction distortion due to inherent errors in multimodal image registration. This approach creatively utilizes the degradation pattern information inherent in PET images as a priori information, replacing the reliance on external anatomical modalities. By identifying and utilizing this degradation pattern information, the reconstruction model can achieve reconstruction quality no less than that of multimodal fusion methods using only single-modal PET data (without the support of CT / MRI multimodal data). This fundamentally avoids registration errors and significantly reduces patient radiation exposure and examination costs. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 The diagram shown is a schematic flowchart of a PET image reconstruction method provided in an embodiment of this specification.

[0041] Figure 2 The diagram shown is a flowchart illustrating the synchronous input of degradation mode information and PET images provided in an embodiment of this specification.

[0042] Figure 3The diagram shown is a flowchart illustrating a training method for a reconstruction model used to reconstruct PET images, as provided in an embodiment of this specification.

[0043] Figure 4 The diagram shown is a schematic representation of an iterative process for the reconstruction model provided in an embodiment of this specification.

[0044] Figure 5 The diagram shown is a structural schematic of a PET image reconstruction apparatus provided in an embodiment of this specification.

[0045] Figure 6 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this specification.

[0046] 501. Acquisition Module;

[0047] 502. Identification module;

[0048] 503. Rebuild module;

[0049] 602. Computer equipment;

[0050] 604. Processing equipment;

[0051] 606. Storage resources;

[0052] 608. Drive system;

[0053] 610. Input / output module;

[0054] 612. Input devices;

[0055] 614. Output devices;

[0056] 616. Presentation equipment;

[0057] 618. Graphical User Interface;

[0058] 620. Network interface;

[0059] 622. Communication link;

[0060] 624. Communication bus. Detailed Implementation

[0061] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the embodiments of this specification.

[0062] It should be noted that the terms "first," "second," etc., in the description, claims, and accompanying drawings of the embodiments herein are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, apparatus, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0063] It should be noted that the acquisition, storage, use, and processing of data in the technical solutions of the embodiments of this specification all comply with the relevant provisions of national laws and regulations.

[0064] It should be noted that in the embodiments of this specification, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used the solution.

[0065] To reconstruct high-quality images from low-dose PET data, existing techniques rely on multimodal data for reconstruction, such as combining prior anatomical information provided by CT or MRI to improve reconstruction quality.

[0066] However, such methods have the following limitations: Firstly, they lead to increased radiation dose and cost. CT modality introduces additional X-ray radiation, while MRI prolongs scan time and increases equipment costs. Secondly, registration complexity is high. Spatial registration errors between multimodal images can easily cause information misalignment, affecting reconstruction accuracy.

[0067] Furthermore, K. Gong et al., in their 2024 paper "PET image denoising based on denoising diffusion probabilistic model" published in the *European Journal of Nuclear Medicine and Molecular Imaging*, proposed a low-dose PET image denoising method based on the diffusion probability model (DDPM). This study simulates the noise distribution of PET images through a Gaussian diffusion process and utilizes the inverse generation capability of the diffusion model to gradually remove noise, achieving an improvement of over 15% in certain metrics compared to traditional methods.

[0068] However, this method is based on the assumption of a single Gaussian noise, which cannot effectively characterize the characteristics of Poisson-Gaussian mixed noise and the physical degradation mechanism of scattering / attenuation artifacts in low-dose PET, resulting in blurred metabolic regions (such as abdominal lesions) and structural distortions (such as lung boundaries) in reconstructed images in ultra-low dose (1 / 100×) scenarios.

[0069] In view of the problems existing in the prior art, this specification provides a method for reconstructing PET images.

[0070] Figure 1 The diagram illustrates a flowchart of a PET image reconstruction method provided in an embodiment of this specification. The process of reconstructing a PET image is described in this figure. The order of steps listed in the embodiment is merely one possible execution order among many and does not represent the only possible execution order. In actual system or device products, the method can be executed sequentially or in parallel according to the embodiment or the accompanying drawings. Specifically, as shown... Figure 1 As shown, the method can be executed by a computer, and the method may include:

[0071] Step S101: Acquire the positron emission tomography (PET) image to be reconstructed.

[0072] The generation of PET images is a complex process that integrates nuclear physics, electronics, and computer science. Its core purpose is to visualize biochemical activities in the human body, such as glucose metabolism, in the form of images.

[0073] The generation of PET images is illustrated below. For example, a tracer containing a radioactive isotope is injected into the subject's body. This tracer participates in the body's physiological metabolic processes. During the decay of the radioactive isotope, it releases positrons. These positrons travel a very short distance in the surrounding tissues before annihilating with electrons, producing photons. Detectors around the subject capture these photons, and after a series of correction processes, a sine wave is generated. A sine wave can be understood as a projection of an object at different projection angles. Further processing based on the sine wave yields a PET image that clinicians can use for diagnosis; its form can be a cross-sectional image, a coronal image, etc.

[0074] However, during the generation of PET images, some noise is inevitably introduced due to factors such as the imaging physical mechanism and the inherent influence of the electronic system. This noise can be understood as a degradation component, which leads to the degradation of image quality and affects the accuracy of diagnosis.

[0075] The PET image reconstruction method provided in this specification aims to identify and remove degenerative components in PET images using a network model.

[0076] In this step, the PET image to be reconstructed is acquired. For example, the PET image to be reconstructed can be obtained from a clinical PET scanning system.

[0077] Step S102: Input the PET image into a pre-trained degradation pattern recognition model to obtain the degradation pattern information of the PET image.

[0078] Traditional PET reconstruction methods or PET reconstruction methods based on deep learning models attempt to process all PET images using a fixed strategy, while ignoring the significant differences in image degradation patterns under different imaging conditions.

[0079] In the embodiments of this specification, before performing actual denoising reconstruction on low-quality PET images, a pre-trained degradation pattern recognition model is first used to identify degradation pattern information of low-quality PET images.

[0080] In some embodiments of this specification, degradation mode information includes physical parameter information related to the imaging process and state information related to image post-processing. The physical parameter information related to the imaging process is determined before the scan begins and directly affects the parameters of the imaging process. Examples include tracer type, dose level, and scan mode.

[0081] The principle by which tracer type affects the imaging process is as follows: different tracers have different biodistribution characteristics. For example, FDG reflects glucose metabolism and is highly uptaken in the brain, myocardium, and most malignant tumors; while PSMA is specifically enriched on the surface of prostate cancer cells. The dose level directly determines the total photon count that the detector can receive, significantly affecting the signal-to-noise ratio and Poisson noise level of the image. The scanning mode also affects the counting statistics, thus affecting the image imaging process.

[0082] Image post-processing related status information includes the processing flow from generating the initial PET image to generating a clinically usable PET image. For example, this includes attenuation correction status, i.e., whether the PET image has been attenuated using an attenuation map acquired using CT, MRI, or radiometric scanning sources.

[0083] In some embodiments of this specification, the degradation mode information specifically includes tracer type, tracer dosage level, and correction status; wherein, the correction status indicates whether the PET image has been corrected using other modal data.

[0084] For example, the process of training a degradation pattern recognition model is described below.

[0085] First, training data is acquired. This data can come from a large number of clinical PET images from multiple centers and various models of PET scanners. These images need to be accompanied by authentic metadata tags. For example, metadata tags include: 1. Tracer type, such as FDG / DOTA / FAPI / PSMA, etc. 2. Tracer dose level, divided into discrete levels based on the ratio of injected activity to standard dose, such as 1 / 10x, 1 / 20x, 1 / 50x, 1 / 100x, etc. 3. Calibration status, a binary label, such as calibrated or uncalibrated.

[0086] Preprocess the training data. For example, standardize the images to eliminate intensity differences caused by different scanning devices. Alternatively, use random cropping, rotation, or flipping to increase data diversity and improve the model's generalization ability and robustness.

[0087] Training data is input into a pre-designed model architecture for training. The backbone network of this architecture can be a lightweight and efficient network (e.g., MobileNetV2, ShuffleNet) serving as a feature extractor. Additionally, depthwise separable convolutions and attention mechanisms can be introduced to capture the global contextual dependencies of the image while maintaining computational efficiency. Specifically, depthwise separable convolutions reduce the number of parameters, while the attention mechanism enhances the feature representation of key regions, such as high-uptake regions of specific tracers, through multi-head interactions. After feature extraction by the backbone network, multiple parallel fully connected layers can be connected to predict tracer type, tracer dose level, and correction state, respectively. For example, the extracted features are mapped to a classification space via fully connected layers, outputting three labels: tracer type (four-class classification: FDG / DOTA / FAPI / PSMA), dose level (four-class classification: 1 / 10× / 1 / 20× / 1 / 50× / 1 / 100×), and correction state (two-class classification: corrected / uncorrected).

[0088] Experiments show that the trained degradation pattern recognition model performs excellently in classification tasks: the dose classification accuracy is 97.59%, the tracer classification accuracy is 99.59%, and the average accuracy of the corrected state classification on the four types of tracers is 97.57%.

[0089] Step S103: Input the PET image and the degradation pattern information into a pre-trained reconstruction model to obtain a reconstructed PET image; wherein the reconstruction model uses the degradation pattern information as prior information to predict and remove the degradation components of the PET image.

[0090] In the embodiments described in this specification, degradation mode information of the PET image to be reconstructed is identified, and the degradation mode information and the PET image to be reconstructed are simultaneously input into the reconstruction model. This provides targeted prior guidance for the reconstruction model, thereby fully exploring the implicit global and local features in the PET image, significantly improving the signal-to-noise ratio and diagnostic reliability of low-dose PET images, and ultimately obtaining high-quality and highly stable reconstruction results. Furthermore, it solves the technical problem of poor dynamic adaptability of mainstream models to tracer types and dosage levels, requiring independent training for different scenarios. This promotes the development of low-dose PET imaging technology towards a safer and more intelligent direction.

[0091] Furthermore, existing reconstruction methods heavily rely on CT or MRI images to provide prior anatomical information to aid reconstruction. This not only introduces additional radiation doses, scan time, and costs, but can also lead to reconstruction distortion due to inherent errors in multimodal image registration. This approach creatively utilizes the degradation pattern information inherent in PET images as a priori information, replacing the reliance on external anatomical modalities. By identifying and utilizing this degradation pattern information, the reconstruction model can achieve reconstruction quality no less than that of multimodal fusion methods using only single-modal PET data (without the support of CT / MRI multimodal data). This fundamentally avoids registration errors and significantly reduces patient radiation exposure and examination costs.

[0092] In some embodiments of this specification, see Figure 2 The input of PET images and degradation pattern information into the pre-trained reconstruction model further includes:

[0093] Step S201: Encode the degradation pattern information into a semantic feature vector;

[0094] Step S202: Input the semantic feature vector and PET image into the reconstruction model; wherein, the reconstruction model adjusts the spatial attention weight and / or channel attention weight of its attention mechanism according to the semantic feature vector.

[0095] For example, a pre-trained text encoder, such as the MedCLIP encoder, can be introduced. The text encoder converts degradation mode information into text prompts (Promats). For example, for an uncorrected image of 1 / 50× dose DOTA, the prompt would be "1 / 50× low-dose uncorrected PET image using DOTA tracer".

[0096] Subsequently, the text prompt can be input into the text encoder, which outputs a fixed-dimensional, dense semantic feature vector that contains the deep semantics of the input text and can characterize the specific physical imaging scene in which the PET image is located.

[0097] For PET images input to the reconstruction model, preliminary feature extraction can be performed based on the feature extraction network (e.g., backbone network) within the reconstruction model to obtain an initial image feature map. Subsequently, the semantic feature vector serves as a conditional control signal, along with the initial image features... Figure 1 It is included in the core module that follows the input model and contains the attention mechanism.

[0098] For example, the core module of the reconstruction model includes a multi-scale temporal embedding convolutional attention mechanism. Based on the input semantic feature vector, the allocation of its internal attention weights is dynamically adjusted to achieve adaptive reconstruction strategy. Specific modulation mechanisms may include one or more of the following:

[0099] 1. Modulation of spatial attention weights.

[0100] The model can generate a spatial attention weight map based on the semantic feature vector. This weight map is used to recalibrate the importance of each spatial location in the image feature map. For example, when the semantic feature vector indicates that the tracer is of the FDG type, because FDG-type tracers are highly uptaken in the brain, the model will generate higher spatial attention weights in the brain and surrounding areas, thereby guiding the network model to focus more on this key region during processing, such as enhancing its feature representation.

[0101] 2. Modulation of channel attention weights.

[0102] The model can generate a channel attention weight vector based on the semantic feature vector. This weight vector is used to recalibrate the importance of each feature channel in the image feature map. For example, when the semantic feature vector indicates an "uncorrected" state, the model can significantly enhance the weights of feature channels that are sensitive to attenuation artifacts and scattering noise, while potentially suppressing some minor channels, enabling the network model to more effectively extract and process features associated with the current major degradation type.

[0103] In the embodiments of this specification, by encoding degradation mode information into semantic feature vectors, the semantic feature vectors indirectly guide the reconstruction strategy of the model, enabling the model to adaptively focus on key regions and select key features, thereby further improving the image quality of the reconstructed image.

[0104] Traditional PET reconstruction methods fit the denoised image using only a single Gaussian noise distribution, resulting in a lack of modeling ability for complex degradation patterns unique to PET images, such as Poisson noise and systematic scattering / attenuation artifacts. This manifests as residual quantum noise specks (e.g., granular texture in low-count regions) and non-uniform artifacts (e.g., blurred bone boundaries) in the reconstructed image. To address this issue, some embodiments in this specification employ a multi-stage denoising mechanism. Specifically, the reconstruction model incorporates multiple degradation component prediction branches. Using degradation pattern information as prior information, the model predicts different types of degradation components through its multiple degradation component prediction branches.

[0105] Considering the Poisson noise unique to PET images, systematic scattering / attenuation artifacts, and Gaussian noise commonly found in medical images, some embodiments of this specification include multiple degradation component prediction branches, including a branch for predicting Poisson noise components, a branch for predicting Gaussian noise components, and a branch for predicting physical artifact components.

[0106] Each degradation component prediction branch can share a backbone network. The features extracted by the backbone network can be input into three independent neural network branches, each responsible for predicting a degradation component of a specific physical mechanism.

[0107] The following section introduces the three types of degradation components, taking into account the physical imaging mechanism of PET images and the influence of electronic systems.

[0108] Poisson noise is the most significant source of noise in PET images, stemming from the fact that radioactive decay itself is a quantum random event. The nuclear decay of radioactive tracers is random and independent. Within a given time interval, the number of decaying atoms follows a Poisson statistical distribution. A key characteristic of Poisson noise is its positive correlation with pixel intensity. In images, this manifests as higher absolute noise levels in metabolically active regions with high uptake (such as tumors) due to higher average counts, resulting in noticeable granular or speckled textures.

[0109] Gaussian noise typically originates from interference in electronic systems and is the superposition of various electronic noises, such as thermal noise and quantization noise. In images, Gaussian noise is usually spatially relatively uniformly distributed and independent of the image content.

[0110] Physical artifacts are not random noise, but rather unique degradation components inherent in PET image imaging. They mainly include scattering artifacts and attenuation artifacts. Scattering artifacts occur because photons are scattered by atoms within the body before reaching the detector, resulting in energy loss and changes in their direction of motion. This ultimately leads to image blurring and decreased contrast, especially around organ boundaries. Attenuation artifacts occur because photons are absorbed as they pass through human tissue; the higher the density of photons passing through the tissue, the more severe the attenuation. In images without attenuation correction or with inaccurate correction, this manifests as signal distortion and non-uniformity in deep tissues. For example, false bright or dark edges appear at the boundaries between different tissues.

[0111] In the embodiments of this specification, a triple noise estimation framework is designed. Three parallel neural network branches predict Poisson noise, Gaussian noise, and scattering / attenuation degradation components, respectively. The model achieves class-wise suppression of mixed degradation patterns by jointly optimizing the estimation and elimination of these three noise components. The reverse denoising process proposed in this invention overcomes the limitation of traditional reconstruction models that rely solely on Gaussian noise. Through collaborative optimization of multi-task neural network branches, it achieves independent decoupling and targeted elimination of mixed degradation components. This allows the model to preserve key anatomical structures (such as small lesions) while effectively suppressing mixed noise in complex degradation scenarios. For example, in abdominal images, the model can simultaneously eliminate granular artifacts caused by Poisson noise and contrast reduction caused by Gaussian noise, while restoring blood vessel boundaries blurred by scattering effects.

[0112] In some embodiments of this specification, the training sample data for the reconstruction model includes standard PET images and noisy PET images. The noisy PET images are obtained by the following steps: based on a preset total number of time steps T, multiple degradation components corresponding to the physical mechanism of PET imaging are injected in stages through T consecutive noise-adding steps on the standard PET image to generate a continuous image sequence from slightly noisy images to completely noisy images.

[0113] Specifically, the training sample data used to train the reconstruction model is generated through a physical mechanism-guided, phased accumulation noise-adding process. This generation method ensures that the generated noisy images are statistically highly consistent with real low-dose clinical PET images.

[0114] For example, a batch of high-quality standard-dose PET images is acquired. To simulate the degradation process from standard PET images to low-dose PET images, a forward noise-adding process with T noise-adding steps is defined. t is the time step of the current noise-adding step, t=0, 1, 2…T, where t=0 corresponds to a standard PET image, and t=T corresponds to a completely noisy image.

[0115] In each noise addition step, multiple degradation components corresponding to the physical mechanisms of PET imaging are injected. For example, as described above, the degradation components resulting from the physical mechanisms of PET imaging mainly include Poisson noise, Gaussian noise, and physical artifacts. Therefore, to simulate the degradation process from a standard PET image to a low-dose PET image, each noise addition step t involves the simultaneous injection of these three types of physical degradation components.

[0116] For example, for Poisson noise, to address the statistical fluctuations caused by limited photon counting, a Poisson noise component related to pixel intensity is gradually added by analyzing the distribution pattern of metabolic signals in the image, thereby simulating the distribution of Poisson noise. For Gaussian noise, a random noise field following a Gaussian distribution is generated to simulate Gaussian noise caused by electronic systems. Furthermore, systematic degradation offset noise introduced by other factors during the process of transforming a high-quality image into a low-quality image (e.g., non-uniform distortion caused by scattering and attenuation effects) that is difficult to model mathematically directly is added.

[0117] From t=1 to t=T, the above noise addition steps are executed sequentially. For the three noise components, injection intensity coefficients related to time step t can be defined respectively. The three coefficients independently control the accumulation rate of the three types of degradation components as t increases, ensuring that the degradation components of the PET images at each stage in the generated continuous image sequence from slightly noisy image to completely noisy image conform to physical laws.

[0118] Furthermore, for each noise addition step, the cumulative injected noise components that better match the real physical degradation pattern during the process from the standard PET image to the current noisy PET image can be recorded as supervision labels for training the reconstruction model.

[0119] The reconstruction model can then be trained using the constructed image sequence as training data. For example, during training, the model employs the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -5 And through a cosine annealing strategy, it gradually decays to 1×10 over 100 training cycles. -8 This is to avoid model oscillations and improve convergence stability.

[0120] For example, the loss function consists of three independent loss terms. The three types of loss terms are weighted and summed to form a joint optimization objective, which drives the model to learn the distribution characteristics of different noise sources simultaneously during backpropagation and adjust the model parameters.

[0121] In the embodiments described in this specification, a unique noise-adding method is designed based on the physical imaging mechanism of PET images. Training data is generated through a physically driven, staged accumulation noise-adding method, ensuring that the statistical characteristics of noise and artifacts in the noisy images are highly consistent with those of clinical low-dose PET images. Furthermore, three different supervision signals are provided to the model, facilitating the decoupling of the prediction of the three degradation components and accurately estimating each degradation component in mixed degradation. In addition, the generated continuous image sequence includes images at different degradation stages, from slightly noisy to completely noisy, enabling the reconstruction model to systematically learn differentiated reconstruction strategies at different stages along the entire denoising trajectory, significantly improving reconstruction performance.

[0122] For example, comparative experiments were conducted against several existing reconstruction methods (CycleGAN, UNet++, and EMCAD) and the reconstruction method provided in this specification. Experimental results show that the reconstruction method provided in this specification achieves a peak signal-to-noise ratio (PSNR) of 33.39 dB and a structural similarity index (SSIM) of 0.9416 at 1 / 100× dose, significantly outperforming CycleGAN (PSNR 31.07 dB, SSIM 0.7656), UNet++ (PSNR 31.82 dB, SSIM 0.7621), and EMCAD (PSNR 31.01 dB, SSIM 0.8911). In multi-center validation, the model achieved a mean PSNR of 32.72 dB and a mean SSIM of 0.9315 on data from multiple hospitals, representing improvements of 1.34 dB and 0.0889 respectively compared to existing methods (such as RDDM). Visual comparisons show that the reconstruction method provided in this manual outperforms the comparison algorithm in restoring small abdominal lesions and lung boundaries, effectively suppressing artifacts caused by Poisson noise and preserving detailed features of metabolically abnormal areas. Furthermore, the model improves PSNR by 8.48 dB, 6.56 dB, 6.12 dB, and 0.07 dB for different tracers (FDG, PSMA, DOTA, and FAPI), respectively, and SSIM by 0.0245, 0.0499, 0.1272, and 0.02, demonstrating its strong cross-tracer generalization ability.

[0123] See Figure 3 This specification also provides an embodiment of a method for training a reconstruction model for reconstructing PET images, the method comprising the following steps:

[0124] Step S301: Obtain a set of training sample pairs, wherein each training sample pair includes a standard PET image and a noisy image corresponding to different time steps generated by performing a noise-adding step on the standard PET image;

[0125] Step S302: Based on the training sample pair set, iteratively optimize the initialized reconstruction model.

[0126] The construction of the training sample pair set can be found above and will not be repeated here.

[0127] See Figure 4 In each iteration, perform the following steps:

[0128] Step S401: Sample a training sample pair from the training sample pair set, the sample pair containing a standard PET image and a noisy PET image corresponding to the first time step;

[0129] Step S402: Input the noisy PET image corresponding to the first time step, the first time step, and the degradation mode information of the noisy PET image into the reconstruction model to obtain the predicted values ​​of the reconstruction model for multiple degradation components in the noisy PET image;

[0130] Step S403: Calculate the loss function based on the difference between the predicted value and the actual degradation component value corresponding to the first time step;

[0131] Step S404: Update the parameters of the reconstruction model according to the loss function.

[0132] For example, in each training iteration, a training sample pair is randomly selected from the dataset and a time step is randomly chosen. This training sample pair may include I_ and I_t. Where I_ Let I_t represent a standard PET image, and I_t represent a noisy PET image at time step t. The noisy PET image, time step t, and degradation mode information are input into the reconstruction model. The reconstruction model, through its multiple parallel branches, outputs predicted values ​​for different degradation components in the current image. For example, it outputs the predicted Poisson noise component, Gaussian noise component, and physical artifact component. Subsequently, a loss term is calculated based on the difference between the predicted multiple degradation components and the pre-recorded true values ​​corresponding to each degradation component. The Poisson noise loss term, Gaussian noise loss term, and physical artifact loss term are weighted to obtain the total loss. The gradient of the total loss with respect to the model parameters is calculated using the backpropagation algorithm, and the model parameters are updated based on the gradient. Through multiple rounds of iterative training, the model's predicted values ​​are made closer to the true degradation components. The iterative process is repeated until the model loss converges or a preset termination condition is met to complete the training. The trained reconstruction model learns the ability to accurately decouple and predict the three types of degradation components from images in any noisy state.

[0133] The training method for the reconstruction model of PET images provided in the embodiments of this specification, compared with traditional training methods, introduces supervision signals corresponding to three degradation components, and clarifies the training objective as accurately estimating each type of physical degradation component. This allows the model to internalize the physical laws of PET imaging (such as the signal dependence of Poisson noise and the structural nature of artifacts) during the learning process, enabling it to perform mechanism-driven decompositional denoising during inference, rather than blind smoothing, thereby generating high-quality reconstructed images.

[0134] Furthermore, this training method randomly samples time step t during each iteration, meaning the model can learn samples from all degradation stages, from slightly noisy to completely noisy. This training strategy enables the model not only to learn to process images with a fixed noise level, but also to learn denoising strategies for different degradation stages. Therefore, the trained reconstruction model can adaptively process low-quality PET images at various dose levels in clinical settings, exhibiting strong generalization performance.

[0135] Based on the same inventive concept, embodiments of this specification also provide a PET image reconstruction apparatus, such as... Figure 5 As shown, the device includes:

[0136] The acquisition module 501 is used to acquire the positron emission tomography (PET) image to be reconstructed;

[0137] The recognition module 502 is used to input the PET image into a pre-trained degradation pattern recognition model to obtain degradation pattern information of the PET image;

[0138] The reconstruction module 503 is used to input the PET image and the degradation pattern information into a pre-trained reconstruction model to obtain a reconstructed PET image; wherein the reconstruction model uses the degradation pattern information as prior information to predict and remove the degradation components of the PET image.

[0139] The PET image reconstruction apparatus provided in the embodiments of this specification identifies the degradation mode information of the PET image to be reconstructed, and simultaneously inputs the degradation mode information and the PET image to be reconstructed into the reconstruction model. This provides targeted prior guidance for the reconstruction model, thereby fully exploring the implicit global and local features in the PET image, significantly improving the signal-to-noise ratio and diagnostic reliability of low-dose PET images, and ultimately obtaining high-quality and highly stable reconstruction results. It solves the technical problem that mainstream models have poor dynamic adaptability to tracer types and dosage levels, requiring independent training for different scenarios. It significantly improves the generalization ability of the reconstruction model, enhances dynamic adaptability, and reduces model training costs.

[0140] Furthermore, by utilizing the degradation pattern information inherent in PET images as a priori information, the reliance on external anatomical modalities is replaced. By identifying and utilizing this degradation pattern information, the reconstructed model can achieve reconstruction quality no less than that of multimodal fusion methods using only single-modal PET data (without the support of CT / MRI multimodal data). This fundamentally avoids registration errors and significantly reduces patient radiation exposure and examination costs.

[0141] like Figure 6 The diagram illustrates the structure of a computer device according to an embodiment of this specification. The computer device in this embodiment is capable of executing the methods described in the embodiments of this specification. The computer device 602 may include one or more processing devices 604, such as one or more central processing units (CPUs), each of which may implement one or more hardware threads. The computer device 602 may also include any storage resource 606 for storing information of any kind, such as code, settings, data, etc. Without limitation, for example, the storage resource 606 may include any type of RAM, any type of ROM, flash memory, hard disk, optical disk, etc. More generally, any storage resource can use any technology to store information. Furthermore, any storage resource may provide volatile or non-volatile retention of information. Furthermore, any storage resource may represent a fixed or removable component of the computer device 602. In one case, when the processing device 604 executes associated instructions stored in any storage resource or combination of storage resources, the computer device 602 may perform any operation of the associated instructions. The computer device 602 also includes one or more drive systems 608 for interacting with any storage resources, such as hard disk drive systems, optical disk drive systems, etc.

[0142] Computer device 602 may also include an input / output module 610 (I / O) for receiving various inputs (via input device 612) and providing various outputs (via output device 614). A specific output mechanism may include a presentation device 616 and an associated graphical user interface (GUI) 618. In other embodiments, the input / output module 610 (I / O), input device 612, and output device 614 may be omitted, and the device may function solely as a computer device within a network. Computer device 602 may also include one or more network interfaces 620 for exchanging data with other devices via one or more communication links 622. One or more communication buses 624 couple the components described above together.

[0143] Communication link 622 can be implemented in any way, such as via a local area network, a wide area network (e.g., the Internet), a point-to-point connection, or any combination thereof. Communication link 622 may include any combination of hardwired links, wireless links, routers, gateway functions, name servers, etc., governed by any protocol or combination of protocols.

[0144] This specification also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0145] This specification also provides computer-readable instructions, wherein when a processor executes the instructions, the program therein causes the processor to perform the above-described method.

[0146] It should be understood that in the various embodiments of this specification, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this specification.

[0147] It should also be understood that, in the embodiments of this specification, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this specification generally indicates that the preceding and following related objects have an "or" relationship.

[0148] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this specification can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this specification.

[0149] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0150] In the several embodiments provided in this specification, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, devices, or units, or they may be electrical, mechanical, or other forms of connection.

[0151] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments described in this specification, depending on actual needs.

[0152] Furthermore, the functional units in the various embodiments of this specification can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0153] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this specification, in essence, or the part that contributes to the prior art, or all or 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 this specification. 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.

[0154] This specification uses specific embodiments to illustrate the principles and implementation methods of this specification. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this specification. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this specification. Therefore, the content of this specification should not be construed as a limitation of this specification.

Claims

1. A method for reconstructing PET images, characterized in that, The method includes: Acquire positron emission tomography (PET) images to be reconstructed; The PET image is input into a pre-trained degradation pattern recognition model to obtain degradation pattern information of the PET image; The PET image and the degradation pattern information are input into a pre-trained reconstruction model to obtain a reconstructed PET image; wherein the reconstruction model uses the degradation pattern information as prior information to predict and remove the degradation components of the PET image.

2. The method according to claim 1, characterized in that, The degradation mode information includes physical parameter information related to the imaging process and state information related to image post-processing.

3. The method according to claim 1 or 2, characterized in that, The degradation mode information includes tracer type, tracer dosage level, and correction status; the correction status indicates whether the PET image has been corrected using other modal data.

4. The method according to claim 1, characterized in that, Inputting the PET image and the degradation pattern information into the pre-trained reconstruction model further includes: encoding the degradation pattern information into a semantic feature vector; and inputting the semantic feature vector and the PET image into the reconstruction model. The reconstruction model adjusts the spatial attention weights and / or channel attention weights of its attention mechanism based on the semantic feature vector.

5. The method according to claim 1, characterized in that, The reconstruction model uses the degradation pattern information as prior information and predicts different types of degradation components through multiple degradation component prediction branches it contains.

6. The method according to claim 5, characterized in that, The multiple degradation component prediction branches include a branch for predicting Poisson noise components, a branch for predicting Gaussian noise components, and a branch for predicting physical artifact components.

7. The method according to claim 1, characterized in that, The training sample data for the reconstruction model includes: standard PET images and noisy PET images; The noisy PET image is obtained by the following steps: Based on the preset total time steps T, on the basis of the standard PET image, multiple degradation components corresponding to the physical mechanism of PET imaging are injected in stages through T consecutive noise addition steps to generate a continuous image sequence from a slightly noisy image to a completely noisy image.

8. The method according to claim 7, characterized in that, The method further includes: In each noise addition step, the following three types of degradation components are injected: Poisson noise component, Gaussian noise component, and physical artifact component.

9. The method according to claim 7 or 8, characterized in that, The method further includes: For each noise addition step, the true value of each type of degradation component injected is recorded during the process from the standard PET image to the current noisy PET image, and used as a supervision label for training the reconstruction model.

10. A method for training a reconstruction model for reconstructing PET images, characterized in that, The method includes: Obtain a set of training sample pairs, wherein each training sample pair includes a standard PET image and a noisy image corresponding to a different time step, generated by performing a noise-adding step on the standard PET image; Based on the training sample pair set, the initialized reconstruction model is iteratively optimized; In each iteration, the following steps are performed: A training sample pair is sampled from the training sample pair set, the sample pair containing a standard PET image and a noisy PET image corresponding to the first time step; The noisy PET image corresponding to the first time step and the first time step are input into the reconstruction model to obtain the predicted values ​​of the reconstruction model for multiple degradation components in the noisy PET image. The loss function is calculated based on the difference between the predicted value and the actual degradation component value corresponding to the first time step; The parameters of the reconstruction model are updated based on the loss function.

11. A device for reconstructing PET images, characterized in that, The device includes: The acquisition module is used to acquire the positron emission tomography (PET) image to be reconstructed. The recognition module is used to input the PET image into a pre-trained degradation pattern recognition model to obtain degradation pattern information of the PET image; The reconstruction module is used to input the PET image and the degradation pattern information into a pre-trained reconstruction model to obtain a reconstructed PET image; wherein the reconstruction model uses the degradation pattern information as prior information to predict and remove the degradation components of the PET image.

12. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 9 or 10.

Images