A method for generating a pet image by fusing a system response matrix and a neural network

Abstract

This invention belongs to the field of medical imaging equipment technology, specifically relating to a PET image generation method that fuses a system response matrix and a neural network. The method includes the following steps: S1: Obtain an initial image; S2: Perform a fusion operation on the initial image to obtain a fused image; wherein the fusion operation process is as follows: S21: Input the initial image of the current round into an MLEM or OSEM algorithm, and output an updated iterative image. S22: Input the initial image into a trained PET image denoising model, and output a denoised image. S23: Multiply the iterative image with the initial image, and then add it to the denoised image to obtain the fused image. S3: Using the fused image of each round as the initial image for the next round, repeatedly execute the fusion operation in step S2, and output the iteratively fused image, which is the desired high-precision PET image. This invention overcomes the shortcomings of limited accuracy in PET systems and the difficulty in eliminating system noise, which leads to the inability to improve the final imaging quality.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging equipment technology, specifically relating to a PET image generation method and system that integrates system response matrix and neural network, as well as a PET image generation device. Background Technology

[0002] Positron emission tomography (PET) is a type of radionuclide imaging that can visualize the metabolism of labeled biomolecules, as well as the activity of receptors and neurotransmitters, in vivo. It is particularly suitable for early diagnosis of diseases without morphological changes, diagnosis of subclinical lesions, and evaluation of treatment efficacy. To achieve molecular metabolic imaging, PET requires labeling essential substances in the body (such as glucose) with short-lived radionuclides (such as 18F) and then injecting them into the body. When these substances form 511 keV gamma photon pairs during metabolic reactions, PET can capture these photons to obtain their distribution within the body. Therefore, PET imaging can provide various biological information such as blood flow, metabolic rate, receptors, and gene expression, providing important diagnostic information for early screening of brain diseases and research into pathogenesis.

[0003] The limited spatial perception capability of PET is currently the main bottleneck in improving PET imaging quality, and this inadequate spatial perception is the underlying reason why the signal-to-noise ratio (SNR) is difficult to improve. Limited by the spatial perception capability of PET, existing solutions are unable to efficiently eliminate the influence of noise information in imaging, making it difficult to improve the SNR and structural detail. If PET imaging systems can shorten imaging process time and improve image quality to achieve finer-grained anomaly presentation while using the lowest possible dose of radiopharmaceuticals to ensure radiation safety, it will be possible to detect previously unseen lesions, providing a basis for early intervention.

[0004] To overcome the aforementioned shortcomings of PET systems, engineers have proposed enhanced PET devices such as PET-MRI and PET-CT. These new devices essentially utilize magnetic resonance imaging or computed tomography (CT) to enhance the imaging capabilities of PET. However, even with auxiliary enhancement of PET images using other systems, the PET signal may still contain some interfering noise. This noise cannot be eliminated in the system response matrix, therefore, PET images generated using conventional image reconstruction algorithms may still contain some noise information, thus reducing the quality of the reconstructed PET image. Summary of the Invention

[0005] To overcome the limitations of PET systems, such as limited accuracy and difficulty in eliminating system noise, which can easily affect the final image quality, this invention provides a PET image generation method, system, and device that integrates system response matrix and neural network.

[0006] This invention is achieved using the following technical solution:

[0007] A PET image generation method integrating system response matrix and neural network is provided, which is used to generate high-precision PET images based on the detection signals of PET equipment, according to MLEM or OSEM algorithms. The PET image generation method integrating system response matrix and neural network provided by this invention includes the following steps:

[0008] S1: Obtain the initial image λ 0 Initial image λ 0 Use a uniform single-value image or the original image reconstructed from the probe signal using the FBP algorithm.

[0009] S2: For the initial image λ 0 Perform fusion operations to obtain the fused image λ. i The fusion operation process is as follows:

[0010] S21: Set the initial image λ of the current round i. i-1 The input is fed into an MLEM or OSEM algorithm that iteratively updates the system response matrix, and the output is the updated iterative image.

[0011] S22: Set the initial image λ of the current round i. i-1 The input is fed into a trained PET image denoising model, and the output is a denoised image.

[0012] S23: Iterate over the image With the initial image λ i-1 After multiplication, then combined with the denoised image. Add them together to obtain the fused image λ for the current round. i ,Right now:

[0013]

[0014] S3: Using the fused image λ obtained in each round respectively i As the initial image for the next round, the fusion operation in step S2 is executed repeatedly, and the fused image λ after multiple iterations is output. n This is the high-precision PET image required.

[0015] As a further improvement of the present invention, in step S22, the PET image denoising model includes a decoder in the image autoencoder branch, a denoising branch in the diffusion network, and a conditional branch.

[0016] The data processing procedure for the PET image noise reduction model is as follows:

[0017] S221: Within the conditional branch, for the input raw PET image λ i-1 Perform convolution processing to obtain feature map y1.

[0018] S221: The feature map y1 is processed three times consecutively using CTF Block to obtain the corresponding feature maps y2, y3 and y4 respectively.

[0019] S223: Connect feature maps y1, y2, y3 and y4 to obtain conditional feature map y5.

[0020] S224: The conditional feature map y5 is introduced into the denoising branch of the diffusion network. The control information corresponding to the conditional feature map y5 is integrated into the intermediate layer of U-net through a cross-attention layer. Then, the denoised feature map s0 is generated step by step in the denoising branch.

[0021] S225: The feature map s0 in the latent space is restored to the PET image using the decoder in the autoencoder branch, which is the denoised image.

[0022] As a further improvement of the present invention, the data processing procedure of CTF Block is as follows:

[0023] First, the original feature maps are processed using both the Transformer module and the convolution module. Then, the feature maps obtained from the two processes are concatenated inversely within the channels. Finally, the concatenated feature maps are processed again by the convolution module.

[0024] As a further improvement to the present invention, the method for constructing the PET image denoising model is as follows:

[0025] S01: Construct a self-encoding branch that includes an encoder and a decoder.

[0026] S02: A latent diffusion model is constructed by connecting an autoencoder branch with a classic diffusion model to learn the noise features of PET images in the latent space.

[0027] In the latent diffusion model, the output of the encoder serves as the input to the diffusion branch of the diffusion model; the output of the denoising branch of the diffusion model serves as the input to the decoder.

[0028] S03: Construct a conditional branch containing a convolutional layer and three CTF blocks; the conditional branch is used to generate a corresponding conditional feature map y5 based on the original PET image.

[0029] S04: Combine the conditional branch with the latent diffusion model to connect the conditional feature map y5 output by the conditional branch with the feature image before each denoising step of the denoising branch in the diffusion model.

[0030] S05: Obtain a large number of high-precision PET images and construct the required raw dataset. Use the raw dataset to perform two-stage training on the network model constructed in the previous step.

[0031] S06: Retain the model parameters of the decoder, denoising branch, and conditional branch in the trained network model to obtain the required PET image denoising model.

[0032] As a further improvement of the present invention, the encoder and decoder in the self-encoding branch adopt the upsampling module and downsampling module in U-net, respectively.

[0033] As a further improvement of the present invention, in step S04, the diffusion model includes a diffusion branch and a noise reduction branch; the diffusion branch is used to superimpose noise information onto the feature map s0 through several forward diffusion modules to obtain a random noise feature map s in the latent space. T ;

[0034] The noise reduction branch is used to process the random noise feature map s T Multiple rounds of noise reduction processing are performed, with the number of rounds being the same as the number of forward diffusion modules in the diffusion branch, thereby obtaining the original feature map with noise removed.

[0035] In each round of noise reduction processing, the noise reduction branch first utilizes a U-net with shared parameters based on the feature map s before noise reduction. t A corresponding noise information ∈θ is predicted; then, the inverse diffusion unit is used to transform the feature map s before noise reduction. t After removing the noise information ∈θ, the denoised feature map s is obtained. t-1 .

[0036] As a further improvement of the present invention, the two-stage training process in step S05 is as follows:

[0037] S051: Acquire a large number of high-precision PET images, and use any method to degrade the high-precision images to obtain the corresponding low-precision images.

[0038] S052: Divide the dataset containing all high-precision images into a first training set and a first test set, and use the two sets to train and test the autoencoder branch separately to complete the first stage of training.

[0039] S053: Divide the dataset containing all low-precision images into a second training set and a second test set, and use the two sets to train and test the network model containing autoencoder branch, diffusion model and conditional branch respectively; complete the training task of the second stage.

[0040] As a further improvement of the present invention, in step S052, the objective function used by the autoencoder branch in the individual training process of the first stage is the variational autoencoder loss function.

[0041] In step S053, the objective function used by the network model during the overall training process in the second stage is:

[0042]

[0043] In the above formula, ∈ θ This represents the neural backbone of the model, i.e., the denoising U-Net model that integrates cross-attention, where ∈ represents the true noise value. t y represents the latent vector obtained from the encoder; t This indicates the output of the conditional branch.

[0044] The present invention also includes a PET image generation system that integrates a system response matrix and a neural network. This system employs the aforementioned method for generating PET images using a fused system response matrix and neural network to generate high-precision PET images based on the detection signals from a PET device. The PET image generation system integrating the fused system response matrix and neural network includes: an initial image generation unit, an image iteration unit, an image denoising unit, and an iterative fusion unit.

[0045] The initial image generation unit is used to take the uniform single-value image or the original image reconstructed by the FBP algorithm from the probe signal as the initial image λ. 0 .

[0046] The image iteration unit employs either the MLEM or OSEM algorithm; it is used to reconstruct the corresponding iterative image based on the probe signal by combining the system response matrix.

[0047] The image denoising unit employs a trained PET image denoising model that includes a decoder, a denoising branch, and a conditional branch. The PET image denoising model is used to denoise the input original PET image λ. i-1 Generate a corresponding denoised image

[0048] The iterative fusion unit is used to process the initial image λ using the image iteration unit and the image denoising unit. 0 Iterative fusion calculations are performed to output the final high-precision PET image λ. n The data processing in each round of the iterative fusion unit includes: processing the iterative images... With the initial image λ i-1 After multiplication, then combined with the denoised image. Add them together to obtain the fused image λ for the current round. i .

[0049] The present invention also includes a PET image generation device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it creates a PET image generation system that fuses the system response matrix and the neural network as described above. Then, based on the detection signal of the input PET device, it generates a corresponding high-precision PET image through the data processing method of the PET image generation method that fuses the system response matrix and the neural network as described above.

[0050] The technical solution provided by this invention has the following beneficial effects:

[0051] This invention presents a PET image generation method that integrates system response matrix and neural network. Building upon existing iterative PET image reconstruction methods, it embeds an iterative algorithm into a deep learning model, achieving both high efficiency and interpretability in PET image reconstruction. The introduction of the deep learning model allows image reconstruction to break free from a single dependence on the probe signal, enabling the acquisition of relatively high-quality PET images even under conditions where long scanning times lead to a decrease in the signal-to-noise ratio of the original signal.

[0052] For specific scenarios involving PET images, this embodiment also designs a deep learning network specifically for denoising PET images, namely the PET image denoising model. This network model is a latent diffusion model with conditional branches, and is obtained through two-stage training. After training, the network model can quickly learn and eliminate noise information in PET images in a finite number of iterations, thereby helping to improve the image quality of PET images generated by traditional iterative reconstruction schemes. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a flowchart illustrating how a traditional PET device reconstructs a corresponding PET image based on a detection signal.

[0055] Figure 2 This is a technical roadmap of the improved PET image reconstruction method in Embodiment 1 of the present invention.

[0056] Figure 3 This is a schematic diagram of a PET device that includes a stereo vision system in Embodiment 1 of the present invention.

[0057] Figure 4 This is a schematic diagram illustrating the principle of block matching estimation of the receptor motion field in Embodiment 1 of the present invention.

[0058] Figure 5 This is a schematic diagram illustrating the principle of secondary screening of LOR by combining the three-dimensional spatial model of the receptor in Embodiment 1 of the present invention.

[0059] Figure 6 This is an example diagram illustrating the construction and application of the spatial attenuation map of brain receptors in Embodiment 1 of the present invention.

[0060] Figure 7 This is a probability distribution diagram showing the LOR detected by this voxel during the construction of the system response matrix. In the diagram, squares represent voxels, and line segments represent the LORs passing through each voxel.

[0061] Figure 8 For simplified calculation Figure 7 A planar coordinate system for the lengths of each line segment.

[0062] Figure 9 This is a technical roadmap for the PET image generation method that integrates the system response matrix and the neural network, as provided in Embodiment 4 of the present invention.

[0063] Figure 10 The diagram shows the overall network architecture of the PET image denoising model during its construction process. The dashed lines represent the network architecture of the PET image denoising model that is ultimately retained after training.

[0064] Figure 11 This is a flowchart of the steps involved in constructing the PET image denoising model in Embodiment 4 of the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] Example 1

[0067] Traditional PET image reconstruction methods in PET equipment, such as Figure 1 As shown, the process generally includes the following steps:

[0068] 1. Acquire the raw probe signal RawData, perform crystal calibration and energy calibration on the data to obtain LMS data.

[0069] 2. Time windows and energy windows are used to perform random coincidence calibration on coincidence events to obtain a set of coincidence events with less noise, namely LMC data.

[0070] 3. Perform DeadTime and normalization on the LMC data to obtain calibrated LMC data.

[0071] 4. Finally, the system matrix is ​​cut using the calibrated LMC data and the original PET image is obtained through iterative reconstruction. The original PET image is then subjected to attenuation calibration and scattering calibration to obtain the final 3D-PET image.

[0072] The technical approach of the improved stereoscopic vision-guided PET image reconstruction method in this embodiment is as follows: Figure 2 As shown. The application process specifically includes the following steps:

[0073] I. Data Collection Phase:

[0074] Construct an enhanced PET device incorporating a stereo vision system, such as Figure 3 As shown, the stereo vision system consists of multiple visible light cameras arranged in a ring around the central cavity of the PET device's detection area. Through installation and debugging, the stereo vision system and the detection area in the PET device are made coaxially distributed; the receptor passes through the detection areas of both the PET device and the stereo vision system, completing the detection.

[0075] The enhanced PET device provided in this embodiment includes both a stereo vision system and a PET device. The stereo vision system primarily enhances the spatial perception capability of the PET device; therefore, in practical applications, both need to undergo time axis calibration and spatial coordinate system calibration. The solution provided in this embodiment uses Iterative Closest Point (ICP) to calibrate the two coordinate systems, specifically including the following steps:

[0076] (a) Different detection sites are preset on a standard phantom, and visible light images and PET data of the phantom are acquired simultaneously using a visible light system and PET equipment.

[0077] (b) Generate a voxel set X based on the detection sites detected in the visible light image, and generate a voxel set P based on the detection sites detected in the PET data:

[0078]

[0079] (c) Find the minimum rotation matrix R and translation vector t that minimizes the following expression:

[0080]

[0081] In the above formula, Rp i Indicates p i Perform a rotation transformation; E(R,t) represents the error between X and P after the rotation and translation transformation.

[0082] The method for finding the minimum rotation matrix R and translation vector t is as follows:

[0083] First, transpose one of the voxel sets X and P and multiply it by the other to obtain matrix W; then, perform SVD decomposition on matrix W to obtain U and V; finally, calculate R and t using the following formula:

[0084]

[0085] In the above formula, u x Denotes the centroid of the voxel set X. u p Denotes the centroid of the voxel set P.

[0086]

[0087] (d) Based on the minimum rotation matrix R and translation vector t, the three-dimensional spatial model established based on the visible light image and the PET data are transformed to the same coordinate system.

[0088] Using the stereo vision system and PET device that have completed time axis synchronization and spatial coordinate calibration in the previous steps, visible light images and PET data of the receptor can be acquired simultaneously. By combining the feature information of the overlapping area of ​​the target object between two images obtained from two adjacent high-resolution cameras in the stereo vision system, the local three-dimensional information of the corresponding part of the receptor can be determined. Finally, based on R and t, the local three-dimensional information of each part of the receptor is transformed to the same coordinate system, thereby obtaining the global three-dimensional information of the receptor, which is then saved as a voxel data file.

[0089] II. Data Preprocessing Stage:

[0090] A three-dimensional spatial model of the receptor was constructed using visible light images, and the receptor's motion field M was estimated based on the spatial variations of the three-dimensional spatial model in the time domain. m Motion estimation can be performed using various methods, such as nonparametric motion model-based motion estimation and parametric model-based motion estimation. Nonparametric motion estimation requires explicit boundary conditions to obtain object motion information, such as optical flow equations and random field methods. Parametric model motion estimation assumes that each image patch undergoes rigid body motion and can be described by a set of parameters. Describing motion with parameters can significantly reduce the overhead of encoding motion information.

[0091] The scheme in this embodiment ultimately selects the block matching method to perform motion field M on the receptor. m The estimation process is as follows:

[0092] like Figure 4 As shown, in the three-dimensional spatial model of the receptor based on visible light images, the target in each frame is divided into many non-overlapping blocks, and it is assumed that the displacement of all pixels in the block is the same. Each block in the current frame is taken as the current block, and the block matching criterion is used to find the block most similar to the current block in a specific search range in the reference frame, that is, the matching block. The relative displacement between the matching block and the current block is the motion vector. The motion field of the receptor can be determined according to the motion vector between the matching block and the current block.

[0093] In this embodiment, the final estimated motion field M of the receptor is... m It can be incorporated into the system response matrix of the PET image reconstruction algorithm, thereby eliminating artifacts in PET images caused by receptor displacement during the scanning phase.

[0094] The working principle of traditional PET equipment is as follows: if two detector units capture the light signal of annihilated gamma photons in very close time intervals, it is considered that the two detector units have captured the same response event, and the line connecting them forms a line of response (LOR). The main task of the subsequent PET image reconstruction process is to calculate the specific location of the annihilated gamma photons on the LOR.

[0095] During PET scans, a large number of annihilation radiation events occur simultaneously. Due to various reasons, two gamma photons arriving at two detectors in the same LOR (Location of Response) may not originate from the same annihilation point. Such coincidence events are called false coincidences, including random coincidences and scattering coincidences. False coincidences cause the reconstructed image to fail to reflect the true location of the annihilation radiation, reducing the spatial resolution of the PET image and leading to incorrect lesion localization. Therefore, existing PET data preprocessing typically uses time windows and energy windows to perform random coincidence calibration on coincidence events, obtaining a set of coincidence events with less noise. However, even after random coincidence calibration, some LORs still do not actually pass through the receptor region, indicating that these LORs are not true coincidence events but rather false coincidence events.

[0096] In the solution provided in this embodiment, such as Figure 5As shown, taking brain scan as an example, the three-dimensional spatial model of the receptor established by the visible light image can be used to further screen all LORs that have undergone random coincidence calibration by combining them with the three-dimensional spatial model to remove LORs corresponding to false coincidence events that have not passed through the receptor space. Therefore, this embodiment can further eliminate noise information contained in the reconstructed data, thereby improving the quality of the reconstructed PET image.

[0097] III. PET Image Reconstruction Stage:

[0098] In this embodiment, since a three-dimensional spatial model of the receptor has already been established using visible light images, the spatial distribution of bones and soft tissues in the receptor can be predicted based on the statistical characteristics of human anatomy, generating a spatial attenuation map (μ-map) of the receptor. This spatial attenuation map is then used to perform attenuation calibration on the reconstructed PET images.

[0099] The method for generating a spatial attenuation map is as follows:

[0100] (i) Establish a three-dimensional spatial model of the receptor based on the visible light images acquired by the stereo vision system.

[0101] (ii) Based on the statistical characteristics of human anatomy, the three-dimensional space of the receptor is divided into multiple local regions according to different classifications of bones, soft tissues, tissue fluids and cavities.

[0102] (iii) Based on the attenuation effect of γ-rays in different tissues, different local regions in the three-dimensional space of the receptor are assigned corresponding correction coefficients.

[0103] (iv) The three-dimensional spatial model of the receptor containing the corresponding correction coefficient is used as the corresponding 3D decay volume, i.e., the spatial decay map.

[0104] Taking the brain as an example, a μ-map of the brain can be constructed based on the head's shape features, which can be used for attenuation correction in subsequent iterative reconstruction to avoid the problem of weak imaging signals in the center of the head. Figure 6 As shown, this embodiment proposes to construct a brain attenuation map through the following steps: First, a three-dimensional contour image of the head (skull image) is created based on a visible light image. Then, the skull image is segmented into three parts: the intracranial brain (soft tissue), the skull (bone), and the extracranial scalp (soft tissue), and the thickness and distribution range of each part are determined. Finally, a 3D attenuation volume, i.e., the required μ-map, is created for the skull image based on the characteristics of different tissues.

[0105] In the reconstruction of PET images, the three-dimensional space containing the receptor is defined as the scan reconstruction region (FOV). Currently, iterative algorithms are commonly used to approximate the image, finding the reconstruction result within the FOV that best approximates the true value. Among these, statistical models accurately describe the physical processes and statistical characteristics of PET imaging, allowing for convenient control of various parameters during reconstruction, making them the standard algorithm in PET image reconstruction. In PET imaging models, it is typically assumed that the number of γ-photons arriving at each detector unit follows a Poisson distribution. Let λ be the number of radioactive decays in the j-th voxel of the object per unit time. j It is a Poisson distributed random variable with a mean of . The detector then observes a certain number (N) of particles from voxel j. j =N j The average value is The probability of the number of decays can be expressed as:

[0106]

[0107] In the above formula, N j The variable representing the number of radioactive decays of the j-th voxel.

[0108] Therefore, a detector unit i detects an average value of A linear combination of measurements of decay times from N voxels They also follow a Poisson distribution and are statistically independent. Where, a ij This can represent the probability that the radioactive decay event of the j-th voxel is detected by detector unit i; a ij ∈A.

[0109] At this point, we call A the system response matrix. Therefore, we have:

[0110]

[0111] In the above formula, N represents the number of voxels corresponding to the receptor space; M represents the number of detector units (i.e., the number of LORs); for ease of calculation of the above likelihood function, the natural logarithm is often taken:

[0112]

[0113] Therefore, the λ that maximizes L(λ) is the reconstruction result λ. * The objective function for reconstructing PET images can be written as: λ * =arg*max λ≥0 L(λ), the above problem can be solved using the following iterative formula:

[0114]

[0115] In the above formula, k represents the number of iterations.

[0116] Here is the initial image λ 0 Uniform single-value images or images reconstructed using FBP can be used. The above solution method is called the Maximum Likelihood-Expectation-Maximization (MLEM) algorithm. Based on this, to improve reconstruction speed, the same iterative formula as the MLEM algorithm is chosen, and the entire projected data space is divided into several subsets. Each iteration updates one subset, and the final PET image reconstruction is achieved through parallel processing; this algorithm is called the OSEM algorithm. In the process of reconstructing PET images using the above algorithms, the accurate calculation of the system response matrix A plays a crucial role in improving reconstruction quality.

[0117] In the process of calculating the system response matrix A, such as Figure 7 As shown, a cuboid encompassing the device's field of view (FOV) is divided into several small cubes of equal size, called voxels. When a line segment (LOR) passes through this cuboid, it will pass through a portion of the voxels. The length of the line segment intercepted by a voxel from the LOR (denoted as i) is considered the probability that the LOR was detected by that voxel, which is denoted as 'a' in the iterative formula. ii The system response matrix A can be obtained by calculating the probability that all LORs passing through the FOV are detected by all voxels.

[0118] This embodiment calculates the system response matrix A through the following process: first, an α value is used to represent the length between a point on the ray and the starting point. Figure 8 Taking a single LOR (Level of Rendering) as an example, assuming it is the i-th LOR, there exists a voxel arrangement of M rows and N columns. Then, in the two-dimensional case, the parameter formula is as follows:

[0119] X(α)=P X +α(Q X -P X )

[0120] Y(α)=P Y +α(Q Y - P Y)

[0121] In the above formula, (P X P Y ) are the parametric coordinates of the starting point P, (Q) X Q Y X(α) represents the coordinates of the endpoint Q; X(α) represents the coordinates of the point at a distance α from the starting point in the x-direction; Y(α) represents the coordinates of the point at a distance α from the starting point in the y-direction.

[0122] Therefore, when the values ​​of X(α) and Y(α) are known, the value of α at that point on the ray can be found. The value of α can be obtained by using the inverse transformation of the above formula, as shown in the following formula:

[0123] α=(X(α)-P X ) / (Q X -P X )

[0124] α=(Y(α)-P Y ) / (Q Y -P Y )

[0125] Next, locate the incident and exit points of the ray as it passes through the voxel array. By comparing the α values ​​at the boundaries of the voxel array through which the ray passes, the α values ​​at the incident and exit points of the ray's passage through the voxel array can be determined.

[0126] α X (I)=(I·D X +B X -P X ) / (Q X -P X )

[0127] α Y (J)=(J·D Y +B Y -P Y ) / (Q Y -P Y )

[0128] α min =max{min{α X (M), α X (0)},min{α Y (N), α Y (0)}}

[0129] α max =min{max{α X (M), α X (0)},max{α Y (N), α Y (0)}}

[0130] In the above formula, α X The value of α represents the distance from the point where the ray intersects with a voxel on the X-axis to the starting point; α Y The value of α represents the distance from the intersection of the ray and the voxel on the Y-axis to the starting point; α X (I) represents the value of α after traveling I units in the x-direction; α Y (J) represents the value of α after traveling J units in the y-direction; α X (M) indicates Figure 8 The α value of the top right voxel in the x-direction; α Y (N) represents Figure 8 The top-right voxel's α value in the y-direction; I and J represent the current row and column, with I ranging from 0 to M and J ranging from 0 to N; D X D Y B represents the unit distance in the x and y directions, respectively; X B Y This represents the coordinates of the bottom right corner of the voxel.

[0131] For α min and α max Other α values ​​between these values ​​are categorized into α0 and α1 in this embodiment. X and α Y Two arrays,

[0132] To calculate α X and α Y To obtain these two arrays, we must first calculate which planes the ray passes through. Combined Figure 8 According to the partitioning rules, these planes are equally spaced and parallel. We only need to know the first and last planes the ray passes through along the X and Y axes to find the other planes. The α value of the first and last planes the ray passes through along the X-axis is i. min and i max The calculation formula is as follows:

[0133]

[0134]

[0135] In the above formula, α min and α max These represent the α values ​​of a possible LOR (considered as a ray) passing through the voxel array's incident and exit points, respectively.

[0136] The α value of the ray passing through the first and last planes in the Y-axis direction is j. min and j max The calculation formula is as follows:

[0137]

[0138]

[0139] From the above formula, we can see that the set of planes through which rays pass in the X-axis direction is {i}. min i max The set of planes through which rays pass in the Y-axis direction is {j}. min j max After obtaining the plane through which the ray passes, the array {α} can be obtained.X} and {α Y}:

[0140] {α X}={α X (i min ), α X (i min +1), ...α X (i max )}

[0141] {α Y}={α Y (j min ), α Y (j min +1), ...α Y (j max )}

[0142] Then the array {α X} and {α Y} are merged into a single array {α}, if array {α} X} and {α Y If the same α value appears, only one needs to be kept during the merging process.

[0143] {α}={merge[{α X}, {α Y}])={α(0), ...α(n)}

[0144] The intercept d of the ray through different voxels 12 for:

[0145] l(m)=d 12 [α(m)-α(m-1)], (m=1,...n)

[0146] d 12 =[(P X -Q X ) 2 +(P Y -Q Y ) 2 ] 1 / 2

[0147] In the above formula, l(m) represents the intercept of the current LOR when it passes through the m-th voxel; α(m) represents the α value corresponding to the m-th voxel; and α(m-1) represents the α value corresponding to the (m-1)-th voxel.

[0148] Voxel coordinates through which the ray passes:

[0149]

[0150]

[0151] α mid = [α(m) + α(m-1)] / 2

[0152] In the above formula, α mid The α value represents the position between the (m-1)th voxel and the mth voxel; i m j m This refers to the coordinates of a voxel.

[0153] From i here m j m a can be further calculated ij j = (i m -1)*d+j m Here, d indicates that the system has d rows and d columns of voxels.

[0154] Therefore, the probability a of the current i-th LOR passing through the j-th voxel is... ij for:

[0155]

[0156] From this, the probability of each LOR traversing any voxel can be calculated, ultimately yielding the required system response matrix A, a ii ∈A.

[0157] In the above-described system response matrix generation algorithm, this embodiment further employs a generation algorithm that incorporates receptor space determination to generate the dynamic system response matrix A required for PET image reconstruction; the improved system response matrix A generation algorithm is as follows:

[0158]

[0159] Where, d 12 [α(m)-α(m-1)] represents the distance between the LOR and the intersection points of two consecutive voxels it passes through.

[0160] This embodiment also utilizes the sports field M m The generated system response matrix A is corrected to obtain a corresponding optimization matrix B. m This allows for artifact correction. Traditional methods require independent motion correction for each time point, a time-consuming and computationally expensive process. Therefore, in this embodiment, when the estimated motion amplitude of the motion field is less than a certain set threshold, these time points are grouped into the same motion phase. To translate this into a mathematical expression, two time points are grouped into the same motion phase when the following conditions are met:

[0161] |M m |<β,

[0162] In the above formula, β represents the preset motion field M. m The motion amplitude threshold; sports field M m The subscript m in the equation represents the motion stage, where m = 1, 2, ..., L.

[0163] Using the estimated receptor motion field M m The expected value maximization formula in the iterative algorithm can be adjusted as follows:

[0164]

[0165] in, It is an element in the system response matrix, representing the probability that voxel j is detected along the i-th LOR during motion phase m. Indicates to The accumulation; and These are the average scattering and random estimates along the i-th LOR, respectively; This represents the average sensitivity of voxel j across all phases of motion.

[0166] The acquired sports field M m ∈R J*J By incorporating it into the system response matrix A, the optimization matrix B can be obtained. m :

[0167] B m =ST m AM m m = 1, 2, ..., L

[0168] In the above formula, S is a diagonal matrix representing the correction factor for LOR normalization; T m This is a diagonal matrix used to characterize the motion-dependent correction factor of LOR.

[0169] By using the updated system response matrix after incorporating the influence of the motion field, the iterative algorithm can correct the changes in the detection signal caused by the motion of the receptor during each iteration, thereby eliminating artifacts in the reconstructed PET image.

[0170] Example 2

[0171] This embodiment provides a stereo vision-guided PET image reconstruction system. This system is essentially a data processing device used to implement the method in Embodiment 1. The data processing device adopts the stereo vision-guided PET image reconstruction method as in Embodiment 1 to guide the image reconstruction process under low radioactive tracers in a single PET system using stereo vision information, thereby obtaining high-quality PET images.

[0172] The stereo vision-guided PET image reconstruction system in this embodiment includes: a data acquisition module, a coincidence event generation module, a spatial information generation module, an LOR filtering module, a system matrix update module, and an image reconstruction module.

[0173] The data acquisition module is used to synchronize the stereo vision system and PET equipment in time and space, and to acquire spatiotemporally synchronized circumferential visible light images and PET detection signals.

[0174] The coincidence event generation module is used to generate a series of coincidence events and their response lines (LORs) based on the PET detection signal; and to perform random coincidence calibration on all coincidence events using time windows and / or energy windows to obtain a coincidence event set.

[0175] The spatial information generation module includes a spatial modeling unit and a motion field estimation unit. The spatial modeling unit is used to establish a three-dimensional spatial model of the receptor within the detection area based on the circumferential visible light image acquired by the data acquisition module. The motion field estimation unit is used to supervise the spatial motion of the receptor and estimate the receptor's motion field M using a block matching method. m .

[0176] The LOR filtering module compares the LOR of each matching event in the matching event set with the three-dimensional spatial model of the receptor to determine whether the LOR passes through the three-dimensional space of the receptor. If it does, the corresponding matching event is retained; otherwise, the corresponding matching event is deleted from the matching event set.

[0177] The system matrix update module includes a probability adjustment unit and a motion field fusion unit. The probability adjustment unit adds probabilistic constraints during the generation of the system response matrix in the OSEM algorithm. When the j-th voxel is traversed by the i-th LOR, if the voxel is outside the three-dimensional space of the receptor, the corresponding system matrix parameter a is adjusted. ij Set to 0. The motion field fusion unit is used to convert the receptor's motion field M... m Introduced into the system response matrix A generated using dynamic probabilities, the corresponding optimization matrix B is obtained. m .

[0178] The image reconstruction module uses the OSEM algorithm as the data processing tool, combined with the matching event set processed by the LOR filtering module and the optimized matrix B output by the system matrix update module. m The corresponding PET image is reconstructed.

[0179] In a further optimized embodiment, the stereo vision-guided PET image reconstruction system also includes an attenuation calibration module. The attenuation calibration module is used to generate a corresponding spatial attenuation map based on the three-dimensional spatial module of the receptor established by the spatial modeling unit, and finally use the spatial attenuation map to perform local intensity attenuation calibration on the PET image output by the image reconstruction module.

[0180] Example 3

[0181] This embodiment provides a visible light-enhanced PET device, comprising: a PET device, a stereo vision system, and a data processing device. The PET device utilizes any existing single PET system. The stereo vision system consists of multiple high-resolution cameras uniformly distributed circumferentially along the detection space of the single PET device, and the detection spaces of the stereo vision system and the PET device are coaxial.

[0182] The data processing device uses the stereo vision-guided PET image reconstruction system as described in Example 2, and takes the detection signals and visible light images synchronously acquired by the PET device and the stereo vision system as input. It processes the input data using the stereo vision-guided PET image reconstruction method described above, and finally outputs the reconstructed high-definition fine-grained PET image.

[0183] It should be noted that this embodiment provides an enhanced PET device that includes a stereo vision system and a data processing device. However, in other embodiments, the stereo vision system and its corresponding data processing device can be separated and used as a functional kit to upgrade an existing single PET system. That is, any unit or individual that uses the stereo vision system and its corresponding data processing device provided in this embodiment alone or in combination with a single PET system is within the protection scope of this invention.

[0184] Example 4

[0185] Building upon the basic schemes provided in Examples 1-3, which combine PET detection signals and system matrices to iteratively reconstruct PET images using MLEM or OSEM algorithms, this embodiment further provides a more optimized scheme: a PET image generation method that integrates system response matrices and neural networks. This scheme, while retaining the iterative reconstruction portion of traditional MLEM or OSEM algorithms, utilizes a newly designed PET image denoising model to further improve the reconstructed PET image through a dual-drive fusion of system response matrices and neural networks.

[0186] Specifically, such as Figure 9 As shown, the PET image generation method that fuses the system response matrix and neural network provided in this embodiment includes the following steps:

[0187] S1: Obtain the initial image λ 0 Initial image λ 0 Use a uniform single-value image or the original image reconstructed from the probe signal using the FBP algorithm.

[0188] S2: For the initial image λ 0 Perform fusion operations to obtain the fused image λ. i The fusion operation process is as follows:

[0189] S21: Set the initial image λ of the current round i. i-1 The input is fed into an MLEM or OSEM algorithm that iteratively updates the system response matrix, and the output is the updated iterative image.

[0190] S22: Set the initial image λ of the current round i. i-1 The input is fed into a trained PET image denoising model, and the output is a denoised image. The PET image denoising model provided in this embodiment includes a decoder in the image autoencoder branch, a denoising branch in the diffusion network, and a conditional branch.

[0191] The overall network architecture diagram of the PET image denoising model construction process is as follows: Figure 10 As shown in the figure, ξ represents the encoder; η represents the decoder; s t Represents the latent spatial feature map; Conv represents a convolutional layer; Transformer Block represents a Transformer layer; y * This represents the conditional input feature map. Combined with... Figure 10 The data processing procedure for the PET image denoising model is as follows:

[0192] S221: Within the conditional branch, for the input raw PET image λ i-1 Perform convolution processing to obtain feature map y1.

[0193] S221: The feature map y1 is processed three times consecutively using CTF Block to obtain the corresponding feature maps y2, y3, and y4. The data processing procedure of CTF Block is as follows:

[0194] First, the original feature maps are processed using both the Transformer module and the convolution module. Then, the feature maps obtained from the two processes are concatenated inversely within the channels. Finally, the concatenated feature maps are processed again by the convolution module.

[0195] S223: Connect feature maps y1, y2, y3 and y4 to obtain conditional feature map y5.

[0196] S224: The conditional feature map y5 is introduced into the denoising branch of the diffusion network. The control information corresponding to the conditional feature map y5 is integrated into the intermediate layer of U-net through a cross-attention layer. Then, the denoised feature map s0 is generated step by step in the denoising branch.

[0197] In this embodiment, the cross-attention layer is implemented as follows:

[0198]

[0199]

[0200]

[0201]

[0202] in, These are three mapping matrices acquired through learning. Q, K, and V represent the query vector, index vector, and value vector calculated in the attention mechanism. It is an intermediate representation of U-Net.

[0203] S225: The feature map s0 in the latent space is restored to the PET image using the decoder in the autoencoder branch, which is the denoised image.

[0204] S23: Iterate over the image With the initial image λ i-1 After multiplication, then combined with the denoised image. Add them together to obtain the fused image λ for the current round. i ,Right now:

[0205]

[0206] S3: Using the fused image λ obtained in each round respectively i As the initial image for the next round, the fusion operation in step S2 is executed repeatedly, and the fused image λ after n iterations is output. n This is the high-precision PET image required.

[0207] In the PET image generation method that integrates the system response matrix and the neural network in this embodiment, the reconstructed initial image and the reconstructed images obtained in each subsequent iteration are simultaneously processed in both the iterative algorithm and the PET image denoising model. The two are then fused, resulting in a high-precision PET image after several iterations. The iterative algorithm can be any existing algorithm for reconstructing PET images based on the system response matrix, including MLEM or OSEM algorithms, or other feasible algorithms. The PET image denoising network is a novel neural network model designed specifically for the PET image denoising problem in this embodiment.

[0208] Specifically, such as Figure 11 As shown, the construction method of the PET image denoising model designed in this embodiment is as follows:

[0209] S01: Construct an autoencoder branch containing an encoder and a decoder. In practical applications, the encoder and decoder in this embodiment's autoencoder branch respectively adopt the upsampling module and downsampling module from U-net. Of course, in other embodiments, other trained upsampling and downsampling modules can also be selected to construct the autoencoder branch required in this embodiment.

[0210] S02: An autoencoder branch is connected to a classic diffusion model to construct a latent diffusion model that can be used to learn the noise features of PET images in the latent space. In the latent diffusion model, the encoder output serves as the input to the diffusion branch of the diffusion model; the output of the denoising branch of the diffusion model serves as the input to the decoder.

[0211] S03: Construct a conditional branch containing one convolutional layer and three CTF blocks. The conditional branch is used to generate a corresponding conditional feature map y5 based on the original PET image.

[0212] S04: Combine the conditional branch with the latent diffusion model to connect the conditional feature map y5 output by the conditional branch with the feature image before each denoising step of the denoising branch in the diffusion model.

[0213] The diffusion model used in this embodiment includes a diffusion branch and a noise reduction branch; the diffusion branch is used to superimpose noise information onto the feature map s0 through several forward diffusion modules to obtain a random noise feature map s in the latent space. T The noise reduction branch is used to process the random noise feature map s. T Multiple rounds of noise reduction processing are performed, with the number of rounds being the same as the number of forward diffusion modules in the diffusion branch, thereby obtaining the original feature map with noise removed.

[0214] Specifically, in each round of noise reduction processing, the noise reduction branch first utilizes a U-net with shared parameters based on the feature map s before noise reduction. t A corresponding noise information ∈θ is predicted; then, the inverse diffusion unit is used to transform the feature map s before noise reduction. t After removing the noise information ∈θ, the denoised feature map s is obtained. t-1 .

[0215] S05: Obtain a large number of high-resolution PET images to create a raw dataset. Use this raw dataset to perform two-stage training on the network model built in the previous steps. The two-stage training process is as follows:

[0216] S051: Acquire a large number of high-precision PET images, and use any method to degrade the high-precision images to obtain the corresponding low-precision images.

[0217] S052: Divide the dataset containing all high-precision images into a first training set and a first test set, and use the two sets to train and test the autoencoder branch separately to complete the first stage of training.

[0218] The objective function used by the autoencoder branch during the first stage of separate training is the variational autoencoder loss function.

[0219] S053: Divide the dataset containing all low-precision images into a second training set and a second test set, and use the two sets to train and test the network model containing the diffusion model and conditional branch respectively; complete the training task of the second stage.

[0220] The objective function used by the network model during the second stage of overall training is:

[0221]

[0222] In the above formula, ∈ θ This represents the neural backbone of the model, i.e., the denoising U-Net model that integrates cross-attention; ∈ represents the true noise value. ε(x) represents the output of the autoencoder branch encoder; z t y represents the latent vector obtained from the encoder; t Indicates the output of the conditional branch;

[0223] S06: Retain the model parameters of the decoder, denoising branch, and conditional branch in the trained network model to obtain the required PET image denoising model.

[0224] Interpretability and generalizability are essential requirements for PET reconstruction. In Example 1 and the prior art, the model-based iterative reconstruction algorithm can ensure that the generated PET image reflects the true radioactivity distribution. The addition of a PET image denoising model in this embodiment further leverages deep learning to fully utilize existing data and improve reconstruction performance.

[0225] Example 5

[0226] This embodiment provides a PET image generation system that integrates a system response matrix and a neural network. This system is the data processing system used to implement the scheme in Embodiment 4. During operation, the system primarily employs the PET image generation method integrating the system response matrix and neural network from Embodiment 4 to generate high-precision PET images based on the detection signals from the PET device. The PET image generation system integrating the system response matrix and neural network provided in this embodiment includes: an initial image generation unit, an image iteration unit, an image noise reduction unit, and an iterative fusion unit.

[0227] The initial image generation unit is used to take the uniform single-value image or the original image reconstructed by the FBP algorithm from the probe signal as the initial image λ. 0 .

[0228] The image iteration unit employs either the MLEM or OSEM algorithm; it is used to reconstruct the corresponding iterative image based on the probe signal by combining the system response matrix.

[0229] The image denoising unit employs a trained PET image denoising model that includes a decoder, a denoising branch, and a conditional branch. The PET image denoising model is used to denoise the input original PET image λ. i-1 Generate a corresponding denoised image

[0230] The iterative fusion unit is used to process the initial image λ using the image iteration unit and the image denoising unit. 0 Iterative fusion calculations are performed to output the final high-precision PET image λ. n The data processing in each round of the iterative fusion unit includes: processing the iterative images... With the initial image λ i-1 After multiplication, then combined with the denoised image. Add them together to obtain the fused image λ for the current round. i .

[0231] Example 6

[0232] This embodiment also provides a PET image generation device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it creates a PET image generation system that fuses the system response matrix and the neural network as described above. Then, based on the detection signal of the input PET device, it generates a corresponding high-precision PET image through the data processing method of the PET image generation method that fuses the system response matrix and the neural network as described above.

[0233] The PET image generation device provided in this embodiment is essentially a computer device. The computer device can be a smart terminal, tablet computer, laptop computer, desktop computer, rack server, blade server, tower server, or cabinet server (including independent servers or server clusters composed of multiple servers), etc.

[0234] The computer device in this embodiment includes, but is not limited to, a memory and a processor that can communicate with each other via a system bus.

[0235] In this embodiment, the memory (i.e., the readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of a computer device, such as the hard disk or RAM of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the computer device. Of course, the memory can also include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0236] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of a computer device. In this embodiment, the processor is used to run program code stored in memory or process data.

[0237] Example 7

[0238] This embodiment provides a PET image denoising model, which is the product in Embodiment 4. The network architecture of this PET image denoising model is as follows: Figure 10 The portion within the dashed box is shown. The PET image noise reduction model provided in this embodiment specifically employs the following... Figure 11 The construction method shown is designed and trained to obtain the PET image denoising model. This model can be combined with traditional iterative algorithms, applied in each iteration of the reconstruction process to improve the accuracy of the reconstructed PET images. Simultaneously, the PET image denoising model can also be used as a standalone image post-processing tool to post-process PET images reconstructed from any existing device, reducing image noise and improving image quality.

[0239] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating PET images by fusing system response matrix and neural network, characterized in that, It is used to generate high-precision PET images based on the detection signals of PET equipment, based on MLEM or OSEM algorithms; the PET image generation method that fuses the system response matrix and neural network includes the following steps: S1: Obtain the initial image λ 0 The initial image λ 0 Use a uniform single-value image or the original image reconstructed from the probe signal using the FBP algorithm; S2: For the initial image λ 0 Perform fusion operations to obtain the fused image λ. i The fusion operation process is as follows: S21: Set the initial image λ of the current round i. i-1 The input is fed into the MLEM or OSEM algorithm, and the output is the updated iterative image. S22: Set the initial image λ of the current round i. i-1 The input is fed into a trained PET image denoising model, and the output is a denoised image. S23: Iterate over the image With the initial image λ i-1 After multiplication, then combined with the denoised image. Add them together to obtain the fused image λ for the current round. i ,Right now: S3: Using the fused image λ obtained in each round respectively i As the initial image for the next round, the fusion operation in step S2 is executed repeatedly, and the fused image λ after multiple iterations is output. n This is the required high-precision PET image; The PET image denoising model includes a decoder in an image autoencoder branch, a denoising branch in a diffusion network, and a conditional branch. The construction method of the PET image denoising model is as follows: S01: Construct a self-encoding branch that includes an encoder and a decoder; S02: An autoencoder branch is connected to a classic diffusion model to form a latent diffusion model that can be used to learn the noise features of PET images in the latent space. In the latent diffusion model, the output of the encoder serves as the input to the diffusion branch of the diffusion model; the output of the denoising branch of the diffusion model serves as the input to the decoder. S03: Construct a conditional branch containing one convolutional layer and three CTF blocks; the conditional branch is used to generate a corresponding conditional feature map y5 based on the original PET image; S04: Combine the conditional branch with the latent diffusion model to connect the conditional feature map y5 output by the conditional branch with the feature image before each denoising step of the denoising branch in the diffusion model. S05: Obtain a large number of high-precision PET images and construct the required raw dataset. Use the raw dataset to perform two-stage training on the network model constructed in the previous step. S06: Retain the model parameters of the decoder, denoising branch, and conditional branch in the trained network model to obtain the required PET image denoising model.

2. The PET image generation method based on the fusion of system response matrix and neural network as described in claim 1, characterized in that: In step S22, The data processing procedure for the PET image noise reduction model is as follows: S221: Within the conditional branch, for the input raw PET image λ i-1 Perform convolution processing to obtain feature map y1; S222: The feature map y1 is processed three times consecutively using CTF Block to obtain the corresponding feature maps y2, y3 and y4 respectively; S223: Connect feature maps y1, y2, y3 and y4 to obtain conditional feature map y5; S224: The conditional feature map y5 is introduced into the denoising branch of the diffusion network, and the control information corresponding to the conditional feature map y5 is integrated into the intermediate layer of U-net through a cross-attention layer mapping. Then, the denoised feature map s0 is generated step by step in the denoising branch; S225: The feature map s0 in the latent space is restored to the PET image using the decoder in the autoencoder branch, which is the denoised image.

3. The PET image generation method based on the fusion of system response matrix and neural network as described in claim 2, characterized in that: The data processing procedure for the CTF Block is as follows: First, the original feature maps are processed using both the Transformer module and the convolution module. Then, the feature maps obtained from the two processes are concatenated inversely within the channels. Finally, the concatenated feature maps are processed again by the convolution module.

4. The PET image generation method based on the fusion of system response matrix and neural network as described in claim 1, characterized in that: The encoder and decoder in the autoencoder branch adopt the upsampling module and downsampling module in U-net, respectively.

5. The PET image generation method based on the fusion of system response matrix and neural network as described in claim 1, characterized in that: In step S04, the diffusion model includes a diffusion branch and a noise reduction branch; the diffusion branch is used to superimpose noise information onto the feature map s0 through several forward diffusion modules to obtain a random noise feature map s in the latent space. t ; The noise reduction branch is used to process the random noise feature map s t Multiple rounds of noise reduction processing are performed, with the number of rounds being the same as the number of forward diffusion modules in the diffusion branch, thereby obtaining the original feature map with noise removed. In each round of noise reduction processing, the noise reduction branch first utilizes a U-net with shared parameters based on the feature map s before noise reduction. t A corresponding noise information ∈θ is predicted; then, the inverse diffusion unit is used to transform the feature map s before noise reduction. t After removing the noise information ∈θ, the denoised feature map s is obtained. t-1 .

6. The PET image generation method based on the fusion of system response matrix and neural network as described in claim 1, characterized in that: In step S05, the two-stage training process is as follows: S051: Acquire a large number of high-precision PET images, and use any method to degrade the high-precision images to obtain the corresponding low-precision images. S052: Divide the dataset containing all high-precision images into a first training set and a first test set, and use the two sets to train and test the autoencoder branch separately to complete the first stage of training. S053: Divide the dataset containing all low-precision images into a second training set and a second test set, and use the two sets to train and test the network model containing autoencoder branch, diffusion model and conditional branch respectively; complete the training task of the second stage.

7. The PET image generation method based on the fusion of system response matrix and neural network as described in claim 6, characterized in that: In step S052, the objective function used by the autoencoder branch during the individual training process in the first stage is the variational autoencoder loss function; In step S053, the objective function used by the network model during the overall training process in the second stage is: In the above formula, ∈ θ This represents the neural backbone of the model, i.e., the denoising U-Net model that integrates cross-attention, and ∈ represents the true noise value; ε(x) represents the output of the self-encoding branch encoder; z t y represents the latent vector obtained from the encoder; t This indicates the output of the conditional branch.

8. A PET image generation system integrating system response matrix and neural network, characterized in that, It employs the PET image generation method based on the fusion of system response matrix and neural network as described in any one of claims 1-7, to generate high-precision PET images based on the detection signals of the PET equipment; the PET image generation system based on the fusion of system response matrix and neural network includes: The initial image generation unit is used to take a uniform single-value image or the original image reconstructed by the FBP algorithm from the probe signal as the initial image λ. 0 ; The image iteration unit employs either the MLEM or OSEM algorithm; the image iteration unit is used to reconstruct the corresponding iterative image based on the probe signal by combining the system response matrix. The image denoising unit employs a trained PET image denoising model that includes a decoder, a denoising branch, and a conditional branch; the PET image denoising model is used to denoise the input original PET image λ. i-1 Generate a corresponding denoised image An iterative fusion unit is used to process the initial image λ using the image iteration unit and the image denoising unit. 0 Iterative fusion calculations are performed to output the final high-precision PET image λ. n The data processing procedure of each round of the iterative fusion unit includes: processing the iterative images... With the initial image λ i-1 After multiplication, then combined with the denoised image. Add them together to obtain the fused image λ for the current round. i .

9. A PET image generation 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 computer program, it creates a PET image generation system that integrates the system response matrix and neural network as described in claim 8, and then generates a corresponding high-precision PET image based on the input detection signal of the PET device through the data processing method of the PET image generation method that integrates the system response matrix and neural network as described in any one of claims 1-7.

Images