PET motion signal extraction method, storage medium and equipment
Through the combination of deep neural network and PCA, high-level features in PET data are extracted and noise-reduced, which solves the problem of validity of motion signal extraction under low count acquisition, and improves the accuracy and generalization ability of signal extraction.
Patent Information
- Application Number
- CN202411489984.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2025-02-18
AI Technical Summary
In the case of low count acquisition, the statistical data no longer obeys the Gaussian distribution, and the principal components obtained by traditional linear PCA may not be optimal, affecting the effectiveness of motion signal extraction in high-noise data.
Deep neural network is used to extract high-level implicit features in PET acquisition data, and combined with PCA algorithm, we find the feature space base vector matrix and extract the motion signal.
Higher-level features are extracted through deep neural networks, combined with PCA noise reduction and extraction of motion information, the accuracy and generalization ability of motion signal extraction are improved, and the computational complexity is reduced.
Smart Images
Figure CN119477823A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical imaging technology, and in particular to a PET motion signal extraction method, storage medium and device. Background Art
[0002] Positron emission tomography (PET) is one of the most advanced clinical imaging technologies in the field of nuclear medicine. Its basic principle can be summarized as follows: using radioactive nuclides that emit positrons when they decay (such as 18 F. 11 C, etc.) label the essential substances for life activities to generate tracers, which are then injected into the human body. The metabolic status is reflected by the uptake of the radionuclide-labeled tracers by organs and lesions. It plays an important role in the early diagnosis, diagnosis and staging, efficacy evaluation, restaging, primary lesion search, lesion metastasis monitoring and treatment monitoring of tumors, central nervous system diseases and cardiovascular diseases, and has become an indispensable imaging technology in the clinical diagnosis and treatment of such major diseases.
[0003] During the PET scan, data collection at each position often takes several minutes, which is inevitably affected by the complex dynamics of respiratory movement and heart beats. Therefore, the data collected by PET is essentially a comprehensive reflection of the photons emitted by the moving organs or tumors at different transient positions. In the image reconstruction stage, this process is equivalent to the superposition of the imaging information at all these dynamic positions, which leads to a significant reduction in the resolution of PET images. On the one hand, under steady breathing conditions, the average displacement of the diaphragm driven by breathing can reach about 20 mm, while the average movement of the liver region is maintained at about 11 mm. This naturally occurring movement pattern inevitably leads to the blurring of the boundaries of the organ edge and the lesion area, thereby reducing the contrast of the lesion, and posing a significant challenge to the accuracy of the core indicators of PET quantitative analysis, such as the standardized uptake value (SUV), metabolic volume (MTV), and total glycolysis value (TLG) of the lesion. On the other hand, for the highly dynamic structure of the heart, studies have shown that the coronary arteries can undergo significant displacements of 8 to 23 mm during the cardiac cycle, while the additional movement of the heart and coronary arteries affected by respiratory movement ranges from 5 to 9 mm. This complex movement pattern caused by cardiac pulsation not only weakens the contrast of myocardial imaging, but may also cause pulsation artifacts with uneven distribution of myocardial activity uptake, thus interfering with the diagnosis of myocardial function.
[0004] In terms of eliminating the interference of cardiac pulsation and respiratory motion on PET imaging, the current PET acquisition system generally adopts a gating technology strategy, that is, the motion signal of the heart or breath is collected through a specific device, and a gating signal is generated based on this and sent to the PET acquisition control system in real time. After receiving the gating signal, the system classifies and analyzes the collected data according to the different cycle phases of the motion signal to ensure that the relative position of the lesion or target organ is almost static within the specific phase interval of each motion cycle. Subsequently, image reconstruction of the PET data in the same motion phase can significantly reduce the artifacts caused by motion, thereby optimizing the image contrast and improving the accuracy of quantitative measurement parameters. Comprehensive analysis of images at each phase can not only dynamically observe the periodic changes of images in three-dimensional space, but also provide valuable diagnostic basis for clinical practice, which is of immeasurable value for the precise staging of the disease, the formulation of treatment plans, and the precise delineation of radiotherapy targets. In addition, cardiac gating technology also has the ability to evaluate quantitative parameters of various important physiological functions of the heart, such as left ventricular ejection fraction (LVEF), end-diastolic volume (EDV), end-systolic volume (ESV), etc.
[0005] The gating technology used in current PET imaging generally relies on external instruments to collect gating signals. This process is not only complex to operate and susceptible to interference from external factors, but also invisibly increases the risk of radiation exposure for operators and the discomfort of patients. Especially in the practice of cardiac gated imaging, ECG gating technology, as the mainstream, relies on ECG signals as a trigger source to generate gating signals. However, during the examination, patients may be nervous, have unstable breathing, low ambient temperature, muscle tension and other factors, which may cause interference waves to be mixed into the ECG signals, thereby generating erroneous gating signals.
[0006] Given that human respiratory movements are regulated by both the autonomic nervous system and the central nervous system of the brain, their variability and irregularity are particularly significant, and accurate monitoring becomes a major challenge. Clinically, in order to capture respiratory movements, a variety of position detection systems have been tried, including but not limited to infrared capture devices, cameras monitoring body surface movements, spirometers monitoring lung dynamics, and using temperature-sensitive elements on breathing masks to track changes in lung intake temperature. Although these methods are intended to indirectly evaluate respiratory signals through body surface movements, they are complex to operate and there may be deviations between the collected signals and the actual respiratory movements of the lungs.
[0007] To solve the above problems, the relevant technology innovatively introduced the principal component analysis PCA (Principal Component Analysis) technology, which can automatically calculate and extract the patient's respiratory or cardiac signals directly from the raw data of the PET scan, completely getting rid of the dependence on external dedicated gating equipment. This innovative method only requires the patient to undergo a standard PET scan to reconstruct a high-quality gated phase image, effectively reducing the burden of clinical operations, while significantly reducing the adverse effects of respiratory movement and cardiac pulsation on PET image quality and quantitative analysis. However, in the case of low-count acquisition, the statistical data no longer obeys the Gaussian distribution, and the principal components obtained by traditional linear PCA may no longer be optimal, which will obviously affect the effectiveness of motion signal extraction in high-noise data. Summary of the invention
[0008] Technical issues to be solved
[0009] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a PET motion signal extraction method, storage medium and device, which solve the technical problem that in the case of low count acquisition, the statistical data no longer obeys the Gaussian distribution, the principal components obtained by traditional linear PCA may not be optimal, and the effectiveness of motion signal extraction in high-noise data will be affected.
[0010] Technical Solution
[0011] In order to achieve the above object, the main technical solutions adopted by the present invention include:
[0012] In a first aspect, the present invention provides a PET motion signal extraction method, comprising:
[0013] Obtaining PET acquisition data Among them, S is the number of time frames of dynamic acquisition, and N is the dimension of detection data;
[0014] Build deep neural networks;
[0015] Extract high-level latent features from X based on deep neural networks Where M is the dimension size of the extracted features;
[0016] Finding the eigenspace basis vector matrix based on PCA algorithm Where K is the selected feature dimension size, which specifically includes: based on P, linear projection is performed on the decentralized preprocessed Y to obtain the reduced dimension data Z is the result of projecting Y on P; construct the Lagrangian function based on the optimization formula and obtain the eigenvalue formula by taking the derivative of P; solve the eigenvalue formula to obtain the eigenspace basis vector matrix;
[0017] Take the first principal component of the feature space basis vector matrix to get the motion feature: PET motion signal.
[0018] Optionally, extract high-level latent features from X based on deep neural networks include:
[0019] The autoencoder is trained based on the PET data X. After the training is completed, the spatial representation features output by the encoder are taken. As a high-level latent feature, where M represents the dimension of the feature;
[0020] Or, use an existing pre-trained neural network, input the PET acquisition data X into the pre-trained neural network, and take the result of the last convolution output of the pre-trained neural network As a high-level latent feature, where M represents the dimension size of the extracted feature.
[0021] Optionally, the optimization formula is:
[0022]
[0023] sP T P=1
[0024] Where T is the matrix transpose operation.
[0025] Optionally, the eigenvalue formula is:
[0026]
[0027] where Σ is the covariance matrix of Y.
[0028] Optionally, the feature space basis vector matrix is Corresponding to the eigenvalues λ1≥λ2≥…,λ K .
[0029] Optionally, the method further includes: preprocessing the PET acquisition data.
[0030] Optionally, the preprocessing includes decentralized preprocessing or z-score normalization preprocessing.
[0031] Optionally, decentralized preprocessing is performed according to the following formula 1≤i≤N.
[0032] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed, the method for extracting PET motion signals as described in any one of the first aspects above is implemented.
[0033] In a third aspect, the present invention provides a storage device including a storage medium and a processor, wherein the storage medium stores a computer program, and when the program is executed by the processor, a PET motion signal extraction method as described in any one of the first aspects is implemented.
[0034] Beneficial Effects
[0035] The beneficial effects of the present invention are as follows: a PET motion signal extraction method of the present invention combines the architecture of a deep neural network and a PCA method to enhance the extraction capability of PET motion signals, the deep neural network extracts higher-level implicit features, and the PCA method is used for noise reduction and motion information extraction. Compared with the related art, the method proposed in the present application for extracting motion signals from collected data by combining a deep neural network and a PCA method framework can utilize the powerful feature extraction capability of the neural network, has high impedance to noise, fully mines higher-level latent space features from the collected data, reduces the data dimension, and does not require additional preparation of large-scale training data. In addition, the PCA algorithm is combined to analyze the features extracted by the neural network, extract the running signal, and reduce the computational complexity, which can help understand the meaning of the features extracted by the model, increase the interpretability of the neural network, and improve the generalization ability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 A schematic diagram of a flow chart of a PET motion signal extraction method provided by an embodiment of the present invention;
[0037] Figure 2 A schematic diagram of the structure of a feature extraction network provided by an embodiment of the present invention;
[0038] Figure 3 A schematic diagram of an external ECG-gated PET image and quantitative parameters provided by an embodiment of the present invention;
[0039] Figure 4 A schematic diagram of a gated PET image and quantitative parameters for extracting a cardiac beat signal from PET data using the PET motion signal extraction method of the present invention is provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0040] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.
[0041] A PET motion signal extraction method proposed in an embodiment of the present invention combines the architecture of a deep neural network and a PCA method to further enhance the accuracy of PET motion signal extraction. First, the intermediate layer features are extracted from the collected data through an existing pre-trained network, or a convolutional autoencoder is used to perform unsupervised training on the collected data to extract the features output by the encoder, and then the PCA algorithm is used to obtain more representative motion signal features from the features extracted by the neural network. By utilizing the powerful feature extraction capabilities of deep neural networks, higher-level implicit features can be fully mined from the collected data, rather than relying solely on the collected data itself. At the same time, the PCA algorithm is used to remove redundant information and noise in the features, which can improve the generalization ability of the algorithm. By visualizing the feature space after PCA dimensionality reduction, the logic and rules behind the model can be more intuitively understood, which is conducive to enhancing the interpretability of the algorithm.
[0042] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.
[0043] First, refer to Figure 1 This embodiment provides a PET motion signal extraction method, including:
[0044] S1, obtain PET acquisition data X. Among them, S is the number of time frames of dynamic acquisition, and N is the dimension of detection data.
[0045] During the total acquisition time, PET data is divided into S frames according to a fixed unit duration. Data framing needs to take into account both the impact of statistical noise and the retention of time information.
[0046] Optionally, the method further includes: preprocessing the PET acquisition data.
[0047] The collected data X is preprocessed and used as network input to extract feature information.
[0048] Optionally, the preprocessing includes decentralized preprocessing or z-score normalization preprocessing.
[0049] Optionally, decentralized preprocessing is performed according to the following formula 1≤i≤N.
[0050] S2, build a deep neural network.
[0051] S3, extracts high-level latent features Y from X based on deep neural networks. Where M is the dimension of the extracted features.
[0052] Finding the eigenspace basis vector matrix based on PCA algorithm K is the selected feature dimension size, which specifically includes:
[0053] S4, based on P, linearly project the decentralized preprocessed Y to obtain the dimension-reduced data Z. Z is the result of projecting Y onto P.
[0054] S5, construct the Lagrangian function based on the optimization formula and differentiate P to obtain the eigenvalue formula.
[0055] S6, solving the eigenvalue formula to obtain the eigenspace basis vector matrix.
[0056] S7, taking the first principal component of the feature space basis vector matrix to obtain a PET motion signal with a motion feature of Z1.
[0057] Optionally, the optimization formula is:
[0058]
[0059] Where T is the matrix transpose operation.
[0060] PCA is based on the premise of minimal information loss. It transforms a set of correlated variables into a set of linearly uncorrelated variables through orthogonal transformation. The basis vectors of the transformed feature space are called principal components. The eigenvectors corresponding to the minimum eigenvalues are often related to noise. Discarding them can achieve the effect of noise reduction and signal extraction to a certain extent. Therefore, in order to solve the feature space, it is required that the covariance between the feature variables after spatial transformation is 0 and the internal variance is as large as possible.
[0061] The Lagrange multiplier method is used to transform the original function optimization problem into a problem with equality constraints. The Lagrange function is constructed and differentiated with respect to P, which becomes an eigenvalue problem.
[0062] Optionally, the eigenvalue formula is:
[0063]
[0064] where Σ is the covariance matrix of Y.
[0065] Optionally, the feature space basis vector matrix is Corresponding to the eigenvalues λ1≥λ2≥…,λ K .
[0066] Compared with the traditional method of using PCA to extract respiratory signals or cardiac signals from PET data, the method proposed in this paper that combines deep neural networks and PCA methods to extract motion signals from collected data can use the powerful feature extraction capabilities of neural networks, high noise resistance, and fully explore higher-level latent space features from collected data, reducing data dimensions, and does not require additional large-scale training data. In addition, the PCA algorithm is combined to analyze the features extracted by the neural network and extract the running signal, which reduces the computational complexity, can help understand the meaning of the features extracted by the model, increase the interpretability of the neural network, and improve the generalization ability of the algorithm.
[0067] Optionally, extract high-level latent features from X based on deep neural networks include:
[0068] Solution 1: Train the autoencoder based on PET data X; after training, take the spatial representation features output by the encoder As a high-level latent feature, where M represents the dimension size of the feature.
[0069] Or, Option 2: Use an existing pre-trained neural network, input the PET acquisition data X into the pre-trained neural network, and take the result of the last convolution output of the pre-trained neural network As a high-level latent feature, where M represents the dimension size of the extracted feature.
[0070] In the implementation of Solution 1, a convolutional autoencoder can be used as the feature extraction network, and its structure is as follows: Figure 2 As shown. It consists of two parts, the encoder and the decoder. The encoder is constructed using a series of convolutional layers and pooling layers, where the convolutional layer captures spatial features from the input data through convolution operations, and the pooling layer reduces the feature map dimension through maximum pooling to compress the data. The encoder maps the input PET data X to a potential spatial representation feature Y. The decoder has a symmetrical structure to the encoder, except that the original pooling layer is replaced by a deconvolution layer for restoring the image size, and finally the decoder obtains the reconstructed image. The convolutional autoencoder minimizes the input data X and the reconstruction result The reconstruction error is used to optimize the model parameters. The loss function L is defined as follows:
[0071]
[0072] Use convolutional autoencoders to extract features. First, use the data X collected by PET as the network input for unsupervised training, and use the stochastic gradient descent optimization algorithm to optimize the network model parameters until convergence. Then use the trained convolutional autoencoder to extract features from the input data X, and pass X through the encoder part to obtain the feature map Y.
[0073] Solution 2: Use the trained neural network. This network is only used as a feature extraction network. It no longer requires a large amount of PET acquisition data for network training, and can directly extract features from the current acquisition data X. The convolutional layer results before the output layer of the neural network are extracted as high-level potential features, and the M-dimensional feature vector Y is obtained by processing S time frames.
[0074] The actual effect of the present invention is as follows Figure 3 and Figure 4 As shown, Figure 3 For external ECG-gated PET images and quantitative parameters, Figure 4 The invention is used to extract the gated PET image and quantitative parameters of the cardiac beat signal from the PET data. The gated images and quantitative parameters of the two methods are similar, but the invention does not require the complicated operation of the external ECG gating device.
[0075] The following is a further description of a PET motion signal extraction method of the present application in conjunction with a specific embodiment:
[0076] The patient's chest was scanned in a single bed using a PET device. The heart beat signal was extracted and cardiac gating imaging was performed using external gating and data-driven methods. Multi-phase images and quantitative results were compared using external gating as the gold standard. The external gating signal came from the three-lead ECG signal of the patient collected simultaneously, and the data-driven gating signal came from the algorithm proposed in this invention applied to PET acquisition data.
[0077] In the implementation process of the present invention, the data dimension of the collected data X is N=172312. The PET acquisition time is 10 minutes, 0.2s per frame, a total of 3000 frames, that is, S=3000. The PET data acquisition time is 10 minutes, 0.2s per frame, a total of 3000 frames, that is, S=3000.
[0078] The present invention uses a deep neural network to extract implicit features, and the feature dimension size M=4096.
[0079] When the present invention is used, VGG16 is selected as the feature extraction network to directly extract features from the current collected data X. The convolution layer and pooling layer in the VGG16 network structure are used to process S time frames to obtain an M-dimensional feature vector Y.
[0080] The same reconstruction algorithm and parameters were used for the reconstruction of cardiac gated images by external ECG gating and data-driven methods: ordered subsets expectation-maximization (OSEM), 16 phases, 17 subsets, 2 iterations, and image matrix 192x192. The multi-phase reconstructed images of external ECG gating and data-driven methods were input into the Siemens syngo workstation for cardiac function measurement. The software automatically calculated and output cardiac function parameter indicators, and recorded left ventricular ejection fraction (LVEF), end-diastolic volume (EDV), end-systolic volume (ESV), stroke volume (SV), etc.
[0081] In a second aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed, a PET motion signal extraction method as described in any one of the above-mentioned first aspects is implemented.
[0082] In a third aspect, an embodiment of the present invention provides a storage device, including a storage medium and a processor, wherein the storage medium stores a computer program, and when the program is executed by the processor, a PET motion signal extraction method as described in any one of the first aspects is implemented.
[0083] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0084] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention should also include these modifications and variations.
[0085] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.
Claims
1. A PET motion signal extraction method, characterized in that: include: Obtaining PET acquisition data Among them, S is the number of time frames of dynamic acquisition, and N is the dimension of detection data; Build deep neural networks; Extract high-level latent features from X based on deep neural networks Where M is the dimension size of the extracted features; Finding the eigenspace basis vector matrix based on PCA algorithm Where K is the selected feature dimension size, which specifically includes: based on P, linear projection is performed on the decentralized preprocessed Y to obtain the reduced dimension data Z is the result of projecting Y on P; construct the Lagrangian function based on the optimization formula and obtain the eigenvalue formula by taking the derivative of P; solve the eigenvalue formula to obtain the eigenspace basis vector matrix; Take the first principal component of the feature space basis vector matrix to get the motion feature: PET motion signal.
2. A PET motion signal extraction method according to claim 1, characterized in that: Extract high-level latent features from X based on deep neural networks include: The autoencoder is trained based on the PET data X. After the training is completed, the spatial representation features output by the encoder are taken. As a high-level latent feature, where M represents the dimension of the feature; Or, use an existing pre-trained neural network, input the PET acquisition data X into the pre-trained neural network, and take the result of the last convolution output of the pre-trained neural network As a high-level latent feature, where M represents the dimension size of the extracted feature.
3. A PET motion signal extraction method according to claim 2, characterized in that: The optimization formula is: stPTP=1 Where T is the matrix transpose operation.
4. A PET motion signal extraction method according to claim 3, characterized in that: The characteristic value formula is: where Σ is the covariance matrix of Y.
5. A PET motion signal extraction method according to claim 4, characterized in that: The feature space basis vector matrix is Corresponding to the eigenvalues λ1≥λ2≥…,λ K .
6. A PET motion signal extraction method according to claim 5, characterized in that: The method further includes: preprocessing the PET acquisition data.
7. A PET motion signal extraction method according to claim 6, characterized in that: The preprocessing includes decentralized preprocessing or z-score normalization preprocessing.
8. A PET motion signal extraction method according to claim 7, characterized in that: Decentralized preprocessing is performed according to the following formula 9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the PET motion signal extraction method according to any one of claims 1 to 8 is implemented.
10. A storage device comprising a storage medium and a processor, wherein the storage medium stores a computer program, wherein: When the processor executes the computer program, the PET motion signal extraction method described in any one of claims 1 to 8 is implemented.