Dynamic model parametric imaging method of dynamic medical image based on probabilistic graph model

Abstract

The invention discloses a dynamic model parameterization imaging method of a dynamic medical image based on a probabilistic graph model, which comprises the following steps: collecting dynamic medical image data, and extracting a pixel value of a region of interest and time sequence information of the pixel value from the dynamic medical image data; generating an initialized input value of a kinetic model parameter by using a traditional method; inputting the pixel value, the time sequence information and the initialized input value of the kinetic model parameter into a probability graph model, and carrying out machine learning training by introducing a hidden variable label, neighborhood information and prior information to optimize the kinetic model parameter; and taking an output value of the trained probability graph model as a parameter of the kinetic model, and generating a parameterized image. According to the method, the problems that parameter estimation of the dynamic medical image is unsuitable, the accuracy is low, and spatial continuity of parameter distribution is poor can be solved, so that a more accurate parameterized imaging image can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and more specifically, to a method for parameterizing a kinetic model of dynamic medical images based on a probabilistic graphical model. Background Art

[0002] Currently, parameterized imaging of the kinetic model of dynamic medical images plays an extremely important role in the diagnosis of diseases, the assessment of the patient's physical condition, and the prediction and evaluation of the development of diseases. Parameterized imaging refers to a technique of estimating medical image parameters of interest, such as myocardial blood flow in myocardial perfusion images, drug metabolism rate in dynamic PET images, etc., pixel by pixel and imaging at the corresponding positions of the images.

[0003] For traditional parameterized imaging of the kinetic model of dynamic medical images, generally, traditional least squares methods are used to estimate the relevant kinetic model parameters of dynamic medical images. Since these traditional methods do not use or have limited use of spatial information, the spatial continuity of parameterized imaging is very poor, and the gradient change of the values of the parameter images is severe, resulting in an incoherent image with many broken and mutated parts. At the same time, since the mathematical expression of the kinetic model of dynamic medical images is an ordinary differential equation, the solution of parameter estimation is ill-posed, that is, a slight perturbation of a small input will cause a huge change in the output, which makes the estimated values of the parameters extremely inaccurate. Summary of the Invention

[0004] The object of the present invention is to propose a method for parameterizing a kinetic model of dynamic medical images based on a probabilistic graphical model, to solve the problems of ill-posed parameter estimation, low accuracy, and poor spatial continuity of parameter distribution in dynamic medical images, so as to obtain a more accurate parameterized imaging image.

[0005] To achieve the above object, the present invention proposes a method for parameterizing a kinetic model of dynamic medical images based on a probabilistic graphical model, including:

[0006] Collect dynamic medical image data, and extract the pixel values and the temporal information of the pixel values of the region of interest from the dynamic medical image data;

[0007] Use traditional methods to generate initial input values of the kinetic model parameters;

[0008] Input the pixel values, the temporal information, and the initial input values of the kinetic model parameters into a probabilistic graphical model, and perform machine learning training by introducing latent variable labels, neighborhood information, and prior information to optimize the parameters of the kinetic model;

[0009] Use the output values of the trained probabilistic graphical model as the parameters of the kinetic model, and generate a parameterized image.

[0010] Optionally, extracting the pixel values of the region of interest and the temporal information of the pixel values from the dynamic medical image data includes:

[0011] Extracting the region of interest from the dynamic medical image data by using an automatic or manual method;

[0012] Extracting the values of the pixels within the region of interest and the temporal information of each pixel;

[0013] If the kinetic model requires blood pool pixel information, additionally obtaining the value of each blood pool pixel within the region of interest and averaging it to reduce errors.

[0014] Optionally, generating the initial input values of the kinetic model parameters by using traditional methods includes:

[0015] Selecting the corresponding kinetic model according to the data type. The kinetic models include the Fermi model and the 2CXM model for myocardial perfusion magnetic resonance data, and the 2TCM model, the 1TCM model, and the 3TCM model for dynamic PET data. The difference between the calculated value and the observed value of the kinetic model is: Z = X - f(Y|θ), where Z is the loss function of the model, that is, the difference between the model calculated value and the observed value; f is the kinetic model function; X is the pixel value extracted from the medical image, which is a sequence over time, representing the change of the tracer concentration or signal intensity in the tissue over time; Y represents the input data, describing the change of the tracer concentration in the arterial blood pool over time; θ is the target parameter value of the parametric imaging of the kinetic model, which is used to describe the kinetic behavior of the tracer in the tissue, including physiological parameters and model-related parameters;

[0016] Using the least squares method to find the parameter θ value that minimizes the loss function |X - f(Y|θ)| as the initial input value;

[0017] Using the maximum likelihood estimation method to obtain the initial input values of other parameters of the kinetic model. The other parameters include all other parameters that do not require parametric imaging in the kinetic model but are reflected in the algorithm.

[0018] Optionally, inputting the pixel values, the temporal information, and the initial input values into a probabilistic graphical model and performing machine learning training by introducing latent variable labels, neighborhood information, and prior information includes:

[0019] Introducing latent variable labels for each pixel within the region of interest to distinguish different tissue types on the image;

[0020] Using a Markov random field to introduce neighborhood information so that adjacent pixels have similar pixel values or change slopes;

[0021] Pixels with different labels are divided into different groups according to latent variables, and each group of pixels follows a different distribution, forming a finite mixture model;

[0022] According to the application scenario, other types of prior information are added to the probabilistic graphical model, and the other types of prior information include linear transformation, mathematical topology information, and sampling prior information;

[0023] A graphical model is generated according to the dependency relationships of the various modules of the probabilistic graphical model.

[0024] Optionally, inputting the pixel value, the timing information, and the initialization input value into the probabilistic graphical model, and performing machine learning training by introducing latent variable labels, neighborhood information, and prior information, further includes:

[0025] According to the generated graphical model, for each parameter θ that needs to be parameterized for imaging, find its corresponding Markov blanket;

[0026] Use the probability distribution expression corresponding to the Markov blanket nodes of the parameter θ to obtain the conditional probability distribution expression of the parameter θ;

[0027] Use a sampling algorithm or an optimization algorithm to solve for the distribution of the parameter θ according to the conditional probability distribution expression of the parameter θ, or use the conditional probability distribution of the parameter θ to solve for the value of the parameter θ.

[0028] Optionally, the using a sampling algorithm to solve for the distribution of the parameter θ according to the conditional probability distribution expression of the parameter θ includes:

[0029] Use the Markov chain Monte Carlo algorithm to sample according to the conditional probability distribution expression of the parameter θ, extract m samples of the parameter θ, and remove the first n samples affected by the initial value, where n < m;

[0030] Use the remaining samples to form the distribution of the parameter θ.

[0031] Optionally, the using an optimization algorithm to solve for the value of the parameter θ according to the conditional probability distribution of the parameter θ includes:

[0032] Use the expectation-maximization algorithm or the iterated conditional modes algorithm to optimize the conditional probability distribution of the parameter θ to find the value of the parameter θ that maximizes the conditional probability of the parameter θ.

[0033] Optionally, using the output value of the trained probabilistic graphical model as the parameter of the kinetic model and generating a parameterized image includes:

[0034] If a sampling algorithm is used to obtain the distribution of parameter θ, then the median of this distribution is taken as the value of parameter θ for visualization, and a plotting library in a computer programming language is used to display the parameterized image.

[0035] Optionally, using the output value of the trained probabilistic graphical model as the parameter of the kinetic model and generating a parameterized image includes:

[0036] If an optimization algorithm is used to obtain the value of parameter θ, then the value of parameter θ for each pixel is directly visualized, and a plotting library in a computer programming language is used to display the parameterized image;

[0037] Optionally, the dynamic medical image is a myocardial perfusion image or a dynamic PET image.

[0038] The beneficial effects of the present invention are as follows:

[0039] (1) First, the present invention collects dynamic medical image data, extracts the pixel values of the region of interest and their temporal information, then uses traditional methods to generate the initial input values of the kinetic model parameters. After that, the pixel values, temporal information, and the initial input values of the kinetic model parameters are input into the probabilistic graphical model, and machine learning training is carried out by introducing latent variable labels, neighborhood information, and prior information to optimize the parameters of the kinetic model. Finally, the output value of the probabilistic graphical model is used as the parameter of the kinetic model, and the kinetic model uses these parameters to generate a parameterized image. By integrating spatial information, temporal information, and prior information using the probabilistic graphical model, the present method optimizes the parameter estimation of the kinetic model. The kinetic model uses the model parameters optimized by the probabilistic graphical model to generate a parameterized image. Through this collaborative working mode, the method of the present invention can significantly improve the accuracy and spatial continuity of the kinetic model parameterization imaging of dynamic medical images.

[0040] (2) By introducing the probabilistic graphical model and combining the Markov random field to introduce the spatial neighborhood information between pixels, the present invention significantly improves the accuracy of parameter estimation. The probabilistic graphical model can use spatial information to smooth the parameter estimation results and reduce the influence of noise, thereby obtaining more accurate parameter estimation values.

[0041] (3) By introducing spatial dependence, the probabilistic graphical model ensures the spatial continuity of parameter estimation. This spatial continuity not only improves the quality of the image but also makes it more in line with the physiological reality, facilitating further analysis by doctors and researchers.

[0042] (4) By integrating spatial information, temporal information, and prior information using the probabilistic graphical model, the robustness of the model is significantly improved. Even in the case of high noise or incomplete data, the probabilistic graphical model can still provide reliable parameter estimation results, reducing the influence of misclassification and noise.

[0043] (5) The method of the present invention realizes the automation of parameter estimation through probabilistic graphical models and automated machine learning training, reducing manual intervention. At the same time, through optimized algorithms and efficient computational methods, the processing speed is improved, making it applicable to large-scale datasets and clinical applications.

[0044] (6) The method of the present invention has wide applicability, not only applicable to myocardial perfusion imaging, but also can be applied to various dynamic medical images such as positron emission tomography (PET). In addition, the method can be extended to other medical imaging fields, such as contrast-enhanced CT, ultrasound perfusion imaging, etc.

[0045] The system of the present invention has other characteristics and advantages, which will be obvious from the accompanying drawings incorporated herein and the subsequent detailed description, or will be described in detail in the accompanying drawings incorporated herein and the subsequent detailed description, and these accompanying drawings and detailed description are jointly used to explain the specific principles of the present invention. Description of the Drawings

[0046] By describing the exemplary embodiments of the present invention in more detail in conjunction with the accompanying drawings, the above and other objects, features, and advantages of the present invention will become more obvious. In the exemplary embodiments of the present invention, the same reference numerals generally represent the same components.

[0047] Figure 1 The flowchart of a method for parametric imaging of the kinetic model of dynamic medical images based on probabilistic graphical models according to an embodiment of the present invention is shown.

[0048] Figure 2 The schematic diagram of the graphical model of the probabilistic graphical model constructed in an embodiment of the present invention is shown.

[0049] Figure 3 The comparison diagram of the parametric imaging result of the tumor by the method of the present invention, the parametric imaging of the tumor by the traditional nonlinear least squares method, and the real tumor image in an embodiment of the present invention is shown. Detailed Description of the Embodiments

[0050] The present invention will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0051] Embodiment

[0052] As Figure 1As shown in the figure, this embodiment provides a method for parameterizing the kinetic model of dynamic medical images based on a probabilistic graphical model, including:

[0053] S1: Collect dynamic medical image data, and extract the pixel values of the region of interest and the temporal information of the pixel values from the dynamic medical image data;

[0054] In this step, the collected dynamic medical image data can be myocardial perfusion images or dynamic PET images.

[0055] Among them, extracting the pixel values of the region of interest and the temporal information of the pixel values from the dynamic medical image data includes:

[0056] Using automatic or manual methods to extract the region of interest from the dynamic medical image data;

[0057] Extract the values of the pixels within the region of interest and the temporal information of each pixel;

[0058] If the kinetic model requires blood pool pixel information, additionally obtain the value of each blood pool pixel within the region of interest and average it to reduce errors.

[0059] In one example, the region of interest (ROI) can be extracted and labeled using automatic methods (such as thresholding, Gaussian mixture models, convolutional neural networks, etc.) or manual methods (such as manual drawing). Then, extract the values of the pixels within the region of interest, X1(t), X2(t), …, X N (t), where t is the temporal information and N is the number of pixels; since the data is dynamic imaging data, each pixel has corresponding temporal information, that is, t = 1, 2, …, T; at the same time, if the kinetic model requires blood pool pixel information Y(t) (general kinetic quantitative analysis models require blood pool pixel information), then additionally obtain blood pool pixel values Y1(t), Y2(t), …, Y M (t), and average the blood pool pixel information to reduce errors, so Y(t) = (Y1(t) + Y2(t) + … + Y M (t)) / M.

[0060] S2: Use traditional methods to generate initial input values of the kinetic model parameters;

[0061] This step specifically includes:

[0062] Select the corresponding kinetic model according to the data type. The kinetic model can be any kinetic model for dynamic medical images, such as Fermi, 2CXM for myocardial perfusion magnetic resonance data, or 2TCM, 1TCM, 3TCM, etc. for dynamic PET data. The difference between the calculated value and the observed value of the kinetic model is: Z = X - f(Y|θ), where Z is the loss function of the model, that is, the difference between the model calculated value and the observed value; f is the kinetic model function; X is the pixel value extracted from the medical image, which is a sequence over time, representing the change of tracer concentration or signal intensity in the tissue over time; Y represents the input data, describing the change of tracer concentration in the arterial blood pool over time; θ is the target parameter value of the parametric imaging of the kinetic model, which is used to describe the kinetic behavior of the tracer in the tissue, including physiological parameters and model-related parameters.

[0063] Use the least squares method to find the parameter θ value that minimizes the loss function |X - f(Y|θ)| as the initial input value.

[0064] Use the maximum likelihood estimation method to obtain the initial input value of other parameters λ of the kinetic model. The other parameters λ include all other parameters in the kinetic model that do not require parametric imaging but are reflected in the algorithm, such as the latent variable label of the pixel, the mean of different groups, the method of error, etc.

[0065] S3: Input the pixel value, the timing information, and the initial input value of the kinetic model parameters into the probabilistic graphical model, and perform machine learning training by introducing latent variable labels, neighborhood information, and prior information to optimize the parameters of the kinetic model.

[0066] This step S3 specifically includes:

[0067] S31: Establish a probabilistic graphical model based on the corresponding dynamic imaging scenario. The specific process is as follows:

[0068] S311: Introduce latent variable labels for each pixel in the region of interest to distinguish different tissue types on the image.

[0069] Specifically, introduce latent variable labels L1, L2, …, L for each pixel N ; The meaning of this label is different in different application scenarios, but the core is to distinguish different tissue types on the image. For example, in the myocardial perfusion scenario, the label can be divided into normal myocardial tissue, damaged myocardial tissue, etc.; in the dynamic PET scenario, the label can be divided into liver, myocardium, bone marrow, etc.

[0070] S312: Use the Markov random field to introduce neighborhood information to make adjacent pixels have similar pixel values or change slopes.

[0071] Specifically, the neighborhood information is introduced using a Markov random field. Considering that adjacent tissues have similar properties, adjacent pixels have similar pixel values, pixel change slopes, etc. Therefore, the neighborhood information can be introduced as a penalty term using a Markov random field. Generally, a Markov random field can be abstracted into the form, where w is the weight coefficient, and x and y are predefined parameters (such as pixel values, parameter values, etc.) of two adjacent pixels.

[0072] S313: Classify pixels with different labels into different groups according to the latent variable, and each group of pixels follows a different distribution, forming a finite mixture model;

[0073] Specifically, pixels with different labels are classified into different groups according to the latent variable, and pixels in different groups follow different distributions, such that all pixels follow a finite mixture model. The parameters of different distributions are controlled by the latent variable label information. For example, if we use a Gaussian mixture model to represent the characteristics of different pixel groups, the means and variances of the Gaussian distributions corresponding to each group are different.

[0074] S314: Add other types of prior information to the probabilistic graphical model according to the application scenario, where the other types of prior information include linear transformation, mathematical topology information, and sampling prior information;

[0075] Specifically, other types of prior information are added to the model according to different application scenarios. For example, linear transformation, mathematical topology information, sampling prior information, etc.

[0076] S315: Generate a graphical model according to the dependency relationships of the various modules of the probabilistic graphical model.

[0077] Specifically, the various modules of the probabilistic graphical model are represented using a graph (graphical model) according to the dependency relationships. Figure 2 As an example of the probabilistic graphical model for dynamic PET parametric imaging, the circles in the graph represent random variables, the squares represent observed values or prior information, and the direction of the arrows represents the dependency relationships between the various quantities, that is, A pointing to B means that B is conditionally dependent on A. The descriptions of the various parameters are as follows in the table:

[0078]

[0079] S32: Perform statistical inference on the established probabilistic graphical model, and the specific process is as follows:

[0080] S321: According to the generated graphical model, for each parameter θ that requires parametric imaging, find its corresponding Markov blanket;

[0081] Specifically, for the graph model generated according to step S315, for each parameter θ that requires parametric imaging, find its corresponding Markov blanket (i.e., its parent nodes, child nodes, and co-parent nodes; note that when pointing from node A to node B, A is the parent node of B and B is the child node of A).

[0082] S322: Obtain the conditional probability distribution expression of parameter θ using the probability distribution expressions corresponding to the Markov blanket nodes of parameter θ;

[0083] Specifically, obtain the conditional probability distribution expression of parameter θ using the probability distribution expressions corresponding to the Markov blanket nodes of parameter θ (the establishment of these expressions is obtained according to steps S311 - S314).

[0084] S33: Use a sampling algorithm or an optimization algorithm to solve for the distribution of parameter θ according to the conditional probability distribution expression of parameter θ, or use the conditional probability distribution of parameter θ to solve for the value of parameter θ.

[0085] In this step, using a sampling algorithm to solve for the distribution of parameter θ according to the conditional probability distribution expression of parameter θ includes:

[0086] Using the Markov chain Monte Carlo algorithm, sample according to the conditional probability distribution expression of parameter θ, draw m samples of parameter θ, and remove the first n samples affected by the initial value, where n < m; then, use the remaining samples to form the distribution of parameter θ.

[0087] For example, according to the efficiency of the algorithm, draw 2000 - 11000 samples of parameter θ and discard the first 1000 samples affected by the initial value. The remaining 1000 - 10000 samples of parameter θ can form the distribution of parameter θ.

[0088] In this step, using an optimization algorithm to solve for the value of parameter θ according to the conditional probability distribution of parameter θ includes:

[0089] Using the expectation - maximization algorithm or the iterated conditional modes algorithm to optimize the conditional probability distribution of parameter θ to find the value of parameter θ that maximizes the conditional probability of parameter θ.

[0090] S4: Use the output value of the trained probabilistic graph model as the parameter of the kinetic model and generate a parametric image.

[0091] In this step, if a sampling algorithm is used to obtain the distribution of parameter θ, take the median of this distribution as the value of parameter θ for visualization, and use the plotting library in the computer programming language to display the parametric image.

[0092] If the value of parameter θ is obtained using an optimization algorithm, then directly visualize the value of parameter θ for each pixel, and use the plotting library in a computer programming language to display the parameterized image.

[0093] Specifically, in step S33, the distribution or value of parameter θ has been obtained. If the value of parameter θ is obtained using an optimization algorithm, then directly visualize the value of parameter θ for each pixel to output the parameterized imaging. If the distribution of parameter θ is obtained using a sampling algorithm, then take the median of this distribution as the value of parameter θ for visualization to output the parameterized imaging. Specifically, the output value of the probabilistic graphical model (i.e., the optimized parameter θ) is used as the parameter of the kinetic model. After the kinetic model obtains the value of parameter θ for each pixel, assign this value to the corresponding position of the image, and use the plotting library in a computer programming language (such as Matplotlib in Python, ggplot in R, imshow in MATLIB, etc.) to display this image.

[0094] The following is an example generated using simulated data:

[0095] After injecting the radiopharmaceutical into the patient's body, the PET device can obtain the radioactive information of the radionuclide. Taking the most traditional radiopharmaceutical FDG as an example. FDG has a higher uptake in the tumor site, which can be reflected by different kinetic parameters. Both the nonlinear least squares method and the method proposed in the present invention can estimate the kinetic parameters, but their performances vary greatly. Figure 3 For comparing the imaging results of the method of the embodiment of the present invention with the imaging results of the traditional nonlinear least squares method and the true situation of the tumor, where sdPGM is the method of the present invention, LS is the traditional nonlinear least squares method, and Truth is the true situation. It can be seen that the method of the present invention is closer to the true situation compared with the traditional method, and while having higher estimation accuracy, it also has better spatial continuity, almost eliminating a large amount of mosaic-like estimation noise faced by the traditional method estimation. At the same time, the method proposed in the present invention can more accurately reflect the characteristics of the tumor. Specifically, in Figure 3 the nonlinear least squares method cannot reflect the tumor (the highly expressed circle in the lung) in parameters k2, k3, and Vb, while the imaging result of the method of the present invention is the same as the true situation and can reflect the tumor.

[0096] Compared with the prior art, the method of the present invention has the following advantages:

[0097] (1) Improve the accuracy of parameter estimation

[0098] Traditional methods (such as the least squares method, etc.) often ignore the spatial information between pixels when estimating the kinetic model parameters of dynamic medical images, resulting in a low accuracy of parameter estimation, especially in the case of large image noise or uneven contrast agent distribution.

[0099] The method of the present invention significantly improves the accuracy of parameter estimation by introducing a probabilistic graphical model and incorporating the spatial neighborhood information between pixels through a Markov random field. The probabilistic graphical model can utilize the spatial information to smooth the parameter estimation results and reduce the influence of noise, thereby obtaining more accurate parameter estimation values.

[0100] (2) Improve the spatial continuity of the image

[0101] The kinetic parameter images estimated by traditional methods are often discontinuous in space, with a large number of isolated noise points and irregular gradient changes, resulting in an incoherent image that is difficult to use for clinical diagnosis.

[0102] In the present invention, the probabilistic graphical model ensures the spatial continuity of parameter estimation by introducing spatial dependence. This spatial continuity not only improves the quality of the image but also makes it more in line with the physiological reality, facilitating further analysis by doctors and researchers.

[0103] (3) Improve the robustness of the model

[0104] Traditional methods are sensitive to noise and data outliers, which easily lead to bias in parameter estimation, especially in the case of spatially varying noise commonly found in dynamic medical images.

[0105] In the present invention, the probabilistic graphical model significantly improves the robustness of the model by integrating spatial information, temporal information, and prior information. Even in the case of high noise or incomplete data, the probabilistic graphical model can still provide reliable parameter estimation results, reducing the influence of misclassification and noise.

[0106] (4) Automation and efficiency

[0107] Traditional methods often require manual parameter adjustment or complex preprocessing when dealing with complex dynamic medical images, resulting in low efficiency.

[0108] The method of the present invention realizes the automation of parameter estimation through a probabilistic graphical model and automated machine learning training, reducing manual intervention. At the same time, by optimizing the algorithm and using efficient computational methods, the processing speed is improved, making it applicable to large-scale datasets and clinical applications.

[0109] (5) Wide applicability

[0110] Traditional methods are usually designed for specific types of dynamic medical images and are difficult to be directly applied to other types of images or different application scenarios.

[0111] The method proposed by the present invention has wide applicability. It is not only applicable to myocardial perfusion imaging, but also can be applied to various dynamic medical images such as positron emission tomography (PET), dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), etc. In addition, this method can also be extended to other medical imaging fields, such as contrast-enhanced CT, ultrasound imaging perfusion, etc.

[0112] In summary, by introducing a probabilistic graph model and combining spatial information, temporal information, and prior information, the method of the present invention significantly improves the accuracy, spatial continuity, and robustness of the kinetic model parameterization imaging of dynamic medical images, and solves the problems presented by the traditional method for parameterization imaging of dynamic medical images, such as poor spatial continuity, drastic gradient changes in the values of parameter images, resulting in discontinuous images with many broken and mutated parts, and inaccurate parameter estimation. At the same time, it also realizes automation and high efficiency, and has wide applicability, providing more reliable and efficient technical support for medical image analysis and clinical diagnosis.

[0113] The embodiments of the present invention have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A dynamic model parameterized imaging method for dynamic medical images based on a probabilistic graphical model, characterized in that: include: Collect dynamic medical image data, and extract pixel values ​​of a region of interest and time series information of the pixel values ​​from the dynamic medical image data; Generate initialization input values ​​of kinetic model parameters using traditional methods; Inputting the pixel value, the time series information and the initialization input value of the dynamic model parameters into the probabilistic graphical model, and performing machine learning training by introducing latent variable labels, neighborhood information and prior information to optimize the parameters of the dynamic model; The output values ​​of the trained probabilistic graphical model are used as parameters of the dynamics model and a parameterized image is generated.

2. The method according to claim 1, characterized in that: The step of extracting pixel values ​​of a region of interest and time series information of the pixel values ​​from dynamic medical image data includes: Extracting regions of interest from dynamic medical image data using automatic or manual methods; Extract the values ​​of pixels within the region of interest and the timing information of each pixel; If the dynamic model requires blood pool pixel information, each blood pool pixel value in the region of interest is additionally obtained and averaged to reduce the error.

3. The method according to claim 1, characterized in that The method of generating the initialization input values ​​of the kinetic model parameters by using the traditional method includes: A corresponding kinetic model is selected according to the data type, wherein the kinetic model includes a Fermi model and a 2CXM model for myocardial perfusion magnetic resonance data, and a 2TCM model, a 1TCM model and a 3TCM model for dynamic PET data. The difference between the calculated value and the observed value of the kinetic model is: Z=Xf(Y|θ), wherein Z is the loss function of the model, that is, the difference between the calculated value and the observed value of the model; f is the kinetic model function; X is a pixel value extracted from a medical image, which is a sequence over time, indicating the change of the tracer concentration or signal intensity in the tissue over time; Y represents input data, describing the change of the tracer concentration in the arterial blood pool over time; θ is a target parameter value of parametric imaging of the kinetic model, which is used to describe the kinetic behavior of the tracer in the tissue, including physiological parameters and model-related parameters; Use the least squares method to find the parameter θ value that minimizes the loss function |Xf(Y|θ)| as the initialization input value; The maximum likelihood estimation method is used to obtain the initialization input values ​​of other parameters of the kinetic model. The other parameters include all other parameters in the kinetic model that do not require parametric imaging but are reflected in the algorithm.

4. The method according to claim 1, characterized in that: The step of inputting the pixel value, the time series information and the initialization input value into a probability graph model and performing machine learning training by introducing latent variable labels, neighborhood information and prior information includes: Introducing latent variable labels for each pixel in the region of interest to distinguish different tissue types on the image; The neighborhood information is introduced by using Markov random field, so that adjacent pixels have similar pixel values ​​or change slopes; Pixels with different labels are divided into different groups according to latent variables. Each group of pixels obeys a different distribution, forming a finite mixture model. Adding other types of prior information to the probabilistic graphical model according to the application scenario, the other types of prior information include linear transformation, mathematical topology information and sampling prior information; Generate a graphical model based on the dependencies between modules of the probabilistic graphical model.

5. The method according to claim 4, characterized in that The step of inputting the pixel value, the time series information and the initialization input value into the probability graph model and performing machine learning training by introducing latent variable labels, neighborhood information and prior information also includes: According to the generated graphical model, for each parameter θ that needs to be parameterized for imaging, find its corresponding Markov blanket; The conditional probability distribution expression of the parameter θ is obtained by using the probability distribution expression corresponding to the Markov blanket node of the parameter θ; The distribution of parameter θ is solved according to the conditional probability distribution expression of parameter θ using a sampling algorithm or an optimization algorithm, or the value of parameter θ is solved using the conditional probability distribution according to parameter θ.

6. The method according to claim 5, characterized in that The method of using a sampling algorithm to solve the distribution of the parameter θ according to the conditional probability distribution expression of the parameter θ includes: Using the Markov chain Monte Carlo algorithm, we sample m parameters according to the conditional probability distribution expression of the parameter θ, and remove the first n samples affected by the initial value, where n <m; The remaining samples are used to form the distribution of parameter θ.

7. The method according to claim 5, characterized in that The method of using an optimization algorithm to solve the value of the parameter θ according to the conditional probability distribution of the parameter θ includes: The conditional probability distribution of the parameter θ is optimized using an expectation-maximization algorithm or an iterative conditional peak algorithm to find a value of the parameter θ that maximizes the conditional probability of the parameter θ.

8. The method according to claim 5, characterized in that The method uses the output values ​​of the trained probability graph model as parameters of the dynamics model and generates a parameterized image, including: If the distribution of parameter θ is obtained using a sampling algorithm, the median of the distribution is taken as the value of parameter θ for visualization, and the parameterized image is displayed using a drawing library in a computer programming language.

9. The method according to claim 5, characterized in that The method uses the output values ​​of the trained probability graph model as parameters of the dynamics model and generates a parameterized image, including: If the value of the parameter θ is obtained using an optimization algorithm, the parameter θ value of each pixel is directly visualized, and the parameterized image is displayed using a drawing library in a computer programming language.

10. The method according to claim 1, characterized in that The dynamic medical image is a myocardial perfusion image or a dynamic PET image.

Images