Dynamic PET-CT pharmacokinetic fusion modeling method combining mechanism and data

Abstract

The invention relates to a mechanism and data combined dynamic PET-CT pharmacokinetic fusion modeling method, and belongs to the technical field of pharmacokinetic modeling, and the method comprises the steps: obtaining original image data of a historical case, and constructing a training data set according to the original image data; constructing a dynamic pharmacokinetic model based on a mechanism model and a data-driven network, and completing the training of the dynamic pharmacokinetic model according to the training data set; collecting actual image data of the current subject in a period of time; and inputting the actual image data into a trained dynamic pharmacokinetic model to obtain a prediction result. According to the method, the mechanism model and the data driving network are fused, the complete pharmacokinetic process can be accurately predicted only through image data in a short time in the early stage, the problems that a traditional model depends on long-time scanning and is easily interfered by motion artifacts and the like are solved, the parameter analysis efficiency and accuracy are greatly improved, and the method is suitable for clinical rapid diagnosis.

Description

Technical Field

[0001] This application relates to the technical field of pharmacokinetic modeling, and particularly to a dynamic PET-CT pharmacokinetic fusion modeling method that combines mechanism and data. Background Art

[0002] As a core technology in the field of molecular imaging, dynamic PET-CT imaging has irreplaceable value in pharmacokinetic and pharmacodynamic research. This imaging technology uses radioactive tracers to track the distribution, metabolism, and clearance processes of drugs in real time at the in-vivo level, and combines quantitative modeling to analyze pharmacokinetic parameters; this technology breaks through the limitations of traditional in-vitro experiments on static time points, providing key imaging evidence for revealing drug action mechanisms, optimizing dosing regimens, and exploring individualized treatments.

[0003] In the prior art, the invention patent "PET-CT Dynamic Medical Image Intelligent Quantitative Analysis System and Analysis Method" with the publication number CN105868537B proposed an intelligent quantitative analysis system. This system optimized the parameters of the kinetic model through a hybrid intelligent optimization algorithm such as the nested artificial immune network and gradient descent, and generated accurate regions of interest (ROIs) based on the clustering of time-radioactivity curves of kinetic feature distributions; this method greatly improved the efficiency and accuracy of parameter estimation by optimizing model parameters and precisely locking the regions of interest.

[0004] However, the above method still needs to rely on a complete time-radioactivity curve as data support before the construction of the model. Currently, the generation of the time-radioactivity curve requires continuous dynamic PET-CT scanning imaging for about one hour to capture the pharmacokinetic process; during the long scanning process, due to the poor tolerance of children, the elderly, or critically ill subjects, it is difficult to fix their body positions, which easily leads to motion artifacts in the collected images, affecting the quality of the imaging data, and thus reducing the accuracy of subsequent analysis of pharmacokinetic parameters. Summary of the Invention

[0005] In order to improve the problem that the prior art relies on long continuous dynamic scans, which easily leads to motion artifacts interfering with the imaging quality and then reducing the accuracy of pharmacokinetic parameter analysis, this application provides a dynamic PET-CT pharmacokinetic fusion modeling method that combines mechanism and data.

[0006] In a first aspect, this application provides a dynamic PET-CT pharmacokinetic fusion modeling method that combines mechanism and data, adopting the following technical solutions: A dynamic PET-CT pharmacokinetic fusion modeling method that combines mechanism and data, comprising: Obtain all the original imaging data in historical cases, and construct a training dataset based on all the original imaging data; Construct a dynamic pharmacokinetic model based on a mechanism model and a deep learning network, and complete the training of the dynamic pharmacokinetic model according to the training dataset; Collect the actual imaging data of the current subject in the early stage; Input the actual imaging data into the trained dynamic pharmacokinetic model to obtain a prediction result, where the prediction result is a predicted tissue time-activity curve; The constructing a training dataset according to all the original imaging data includes: According to the original imaging data between all dynamic frames in each historical case, obtain a plasma input function and a standard tissue time-activity curve corresponding to each case; the plasma input function is used to reflect the concentration dynamics of the radioactive tracer in the plasma of the subjects in different historical cases, and the standard tissue time-activity curve is used to reflect the concentration dynamics of the radioactive tracer in the region of interest of the subjects in different historical cases; Divide the original imaging data of each historical case into two parts to obtain early imaging data and full-course imaging data; Construct a training dataset based on the early imaging data, full-course imaging data, plasma input function and standard tissue time-activity curve of each historical case.

[0007] By adopting the above technical solution, obtaining the original imaging data of historical cases to construct a training dataset provides a rich sample basis for model training, enabling the model to learn the pharmacokinetic characteristics and laws of different individuals; constructing a physical framework based on a mechanism model (such as a two-compartment model) ensures that the model has clear biophysical significance and interpretability, and can reflect the basic processes such as the transport and metabolism of the tracer in the body; at the same time, introducing a deep learning network (a parameter adjustment network and a residual prediction network), which are respectively used to capture the time-varying nature of the individual physiological state and compensate for the complex physiological mechanisms not covered by the mechanism model, thereby significantly improving the adaptability of the model to individual differences and the dynamic prediction accuracy; In practical applications, the present application only needs to collect the actual imaging data of the current subject in the initial short term (such as 10 minutes) after injecting the radioactive tracer, and then the trained dynamic pharmacokinetic model can predict the complete pharmacokinetic process (such as 60 minutes), which greatly shortens the clinical scanning time, reduces the discomfort of the subject caused by long-term scanning and the generation of motion artifacts, and improves the data quality.

[0008] In a specific feasible implementation scheme, the constructing a training dataset according to the original imaging data includes: Spatially align the original image data between each pair of dynamic frames in each historical case to obtain registered image data after spatial alignment; Extract the signal of the target region in the registered image data as the plasma input function; Extract the original tissue time-activity curve within a preset region of interest; Normalize the original tissue time-activity curve to obtain a standard tissue time-activity curve.

[0009] In a specific feasible implementation, the dynamic PET-CT pharmacokinetic model includes a mechanistic model and several deep learning networks. The deep learning networks include a parameter adjustment network and a residual prediction network; the mechanistic model is the basic physical framework of the dynamic PET-CT pharmacokinetic model; The parameter adjustment network is used to obtain a tissue time-activity curve to be verified according to the plasma input function corresponding to the actual image data and the standard tissue time-activity curve; The residual prediction network is used to obtain the predicted tissue time-activity curve according to the plasma input function corresponding to the actual image data and the standard tissue time-activity curve.

[0010] In a specific feasible implementation, the parameter adjustment network is specifically used to obtain time-varying parameters according to the plasma input function corresponding to the actual image data and the standard tissue time-activity curve, and correct the population-based parameters of the two-compartment differential equation in the mechanistic model according to the time-varying parameters; the parameter adjustment network is also used to drive the mechanistic model to solve the two-compartment differential equation to obtain a tissue time-activity curve to be verified; The residual prediction network is specifically used to obtain a residual correction term according to the plasma input function corresponding to the actual image data and the standard tissue time-activity curve, and perform non-linear compensation on the tissue time-activity curve to be verified according to the residual correction term to obtain the predicted tissue time-activity curve.

[0011] In a specific feasible implementation, the early image data is used to reflect the actual drug distribution dynamics at the beginning stage after the subject is injected with a radioactive tracer; the whole-course image data is used to reflect the actual drug distribution dynamics from the beginning stage to the end stage after the subject is injected with a radioactive tracer; the predicted tissue time-activity curve is used to reflect the drug concentration dynamics obtained by prediction in the region of interest from the beginning stage to the end stage.

[0012] In a specific feasible implementation, constructing a dynamic pharmacokinetic model based on the mechanistic model and deep learning networks, and completing the training of the dynamic pharmacokinetic model according to the training dataset includes: Design a parameter adjustment network, which is used to accurately reflect the actual physiological process; Design a residual prediction network, which is used to compensate for the systematic deviation of the mechanism model; Based on the early image data, the whole-course image data, the plasma input function, and the standard tissue time-activity curve, fuse and train the mechanism model, the parameter adjustment network, and the residual prediction network to obtain a dynamic pharmacokinetic model.

[0013] In a specific feasible implementation, the process of fusing and training the mechanism model, the parameter adjustment network, and the residual prediction network based on the early image data, the whole-course image data, the plasma input function, and the standard tissue time-activity curve to obtain a dynamic pharmacokinetic model includes: Determine the initial population-based parameters of the mechanism model according to the standard tissue time-activity curve corresponding to the whole-course image data to complete pre-training; Fix the population-based parameters and train the parameter adjustment network according to the standard tissue time-activity curve corresponding to the whole-course image data to obtain a time-varying parameter correction term; Fix the population-based parameters and the trained parameter adjustment network, and train the parameter adjustment network according to the standard tissue time-activity curve corresponding to the whole-course image data to obtain a residual correction term; Perform end-to-end fine-tuning on the entire architecture, and jointly optimize the mechanism model, the parameter adjustment network, and the residual prediction network to obtain a trained dynamic pharmacokinetic model.

[0014] In a specific feasible implementation, in the step of determining the population-based parameters of the mechanism model, a set of initial population-based parameters is determined by non-linear least squares fitting.

[0015] In a specific feasible implementation, the dynamic pharmacokinetic model realizes the data unification of the predicted tissue time-activity curve and the predicted observed concentration output by the mechanism model through an observation function, and the observation function is as follows: ; where, represents the predicted observed concentration output by the mechanism model; represents the vascular fraction, which reflects the volume ratio of blood vessels in the tissue and is a dimensionless parameter, ; represents the concentration contribution of the unabsorbed plasma tracer in the tissue blood vessels to the PET signal, which is used to reflect the weight of the tracer in the blood vessels in the PET signal; Represents the volume ratio of the non-vascular part of the tissue; is the independent variable and represents time; is the model parameter vector; represents the unbound radioisotope tracer concentration in the tissue, i.e., the concentration freely diffusing in the extracellular fluid of the tissue cells; represents the plasma input function.

[0016] In a second aspect, the present application provides an intelligent terminal, adopting the following technical solution: An intelligent terminal, characterized in that it includes a memory and a processor. At least one instruction, at least one program, a code set or an instruction set is stored in the memory, and the at least one instruction, at least one program, the code set or the instruction set is loaded and executed by the processor to implement a dynamic PET-CT pharmacokinetic fusion modeling method combining mechanism and data as described in the first aspect.

[0017] In a third aspect, the present application provides a computer-readable storage medium, adopting the following technical solution: A computer-readable storage medium, in which at least one instruction, at least one program, a code set or an instruction set is stored, and the at least one instruction, at least one program, the code set or the instruction set is loaded and executed by a processor to implement a dynamic PET-CT pharmacokinetic fusion modeling method combining mechanism and data as described in the first aspect.

[0018] In summary, the present application includes at least one of the following beneficial technical effects: 1. Compared with the dependence of traditional models on long-time continuous dynamic scans, the present application only requires image data collection for a short period of time in the early stage to reconstruct the complete pharmacokinetic process, significantly shortening the scan time, reducing the difficulty of body position fixation for children, the elderly or critically ill subjects, and reducing the interference of motion artifacts on data quality, thereby effectively improving the efficiency and accuracy of overall parameter analysis; 2. The present application introduces time-varying parameters through a parameter adjustment network to capture the time-varying nature of individual physiological states, breaking through the limitation of "fixed parameters" in traditional models; 3. The present application compensates for complex factors such as non-linear metabolism and spatial heterogeneity noise not covered by the mechanism model through a residual prediction network, making the predicted results output by the model closer to real physiological signals and breaking through the limitation of "simplified assumptions" in traditional models; 4. The present application takes the differential equation of the mechanism model as the physical backbone and the neural network as the dynamic compensation mechanism. The two achieve complementary advantages through end-to-end training, and establish a modeling architecture that is both interpretable and flexible. Description of the Drawings

[0019] Figure 1 The flowchart of the dynamic PET-CT pharmacokinetic fusion modeling method combining mechanism and data provided by an embodiment of the present application is shown.

[0020] Figure 2 The output process diagram for reflecting the prediction result in step S400 provided by an embodiment of the present application is shown.

[0021] Figure 3 It is a schematic structural diagram of an electronic device disclosed in another embodiment of the present application.

[0022] Explanation of reference numerals: 500, electronic device; 501, processor; 502, communication bus; 503, user interface; 504, network interface; 505, memory. Detailed implementation manners

[0023] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments.

[0024] In the description of the embodiments of the present application, words such as "for example" or "for illustration" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "for example" or "for illustration" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, using words such as "for example" or "for illustration" aims to present relevant concepts in a specific manner.

[0025] In the description of the embodiments of the present application, the meaning of the term "a plurality" refers to two or more. For example, a plurality of systems refers to two or more systems, and a plurality of screen terminals refers to two or more screen terminals. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the technical features indicated. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. The terms "include", "comprise", "have" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0026] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments.

[0027] Existing pharmacokinetic modeling methods mainly adopt compartment models. A compartment model is a mathematical model used to describe the distribution, transport, and metabolism of substances (such as drugs, hormones, toxins, etc.) in a living organism or system. Its core idea is to abstract a complex living organism or system into a combination of several "compartments", and by analyzing the transport rates of substances between compartments, establish kinetic equations to simulate the actual process.

[0028] However, traditional compartment models have significant limitations: The mechanisms of such models are often based on assumptions that simplify the actual physiological situation, that is, usually using preset fixed parameters (such as constant drug transport rates, metabolic constants, etc.) or default parameter changes conforming to linear kinetics, which is essentially different from complex biological systems; On the one hand, the assumption of fixed parameters is difficult to capture the different physiological dynamic characteristics of each subject; in the real physiological environment, parameters such as drug transport rates (such as the activity of transmembrane transport proteins) and metabolic enzyme concentrations (such as the expression level of cytochrome P450) have significant time-varying and individual heterogeneity. For example, the high expression of vascular endothelial growth factor (VEGF) in tumor tissues will cause the drug permeability (transport parameter in the compartment model) to increase dynamically over time, while the static parameters of traditional models cannot depict such changes; On the other hand, the default of linear kinetics will lead to the neglect of spatial heterogeneity and non-linear processes; due to the significant differences in the physiological microenvironments of different regions of the organism (such as the liver, kidney, tumor), drug metabolism may exhibit non-linear characteristics (such as enzyme saturation effects), but compartment models usually regard the entire tissue as a homogeneous space and describe the transport process with linear differential equations, resulting in the model being unable to accurately reflect the true pharmacokinetic behavior (such as the distribution differences of drugs at the tumor margin and center).

[0029] More critically, the fixed parameter model has a high dependence on data collection. Its subsequent parameter estimation process highly depends on the complete tissue time-activity curve (such as 60-minute continuous dynamic PET-CT scan data), and is extremely sensitive to the sampling scheme: high-frequency sampling is required in the early stage to capture the dynamic of the blood flow phase, and extended collection is required in the later stage to accurately fit the terminal trend of metabolism; however, in the actual clinical scenario, due to equipment resolution limitations, counting statistical noise, motion artifacts of subjects (especially children, the elderly, or critically ill subjects), and individual physiological differences, the signal-to-noise ratio of the tissue time-activity curve obtained through long-term collection and analysis is often low, making the estimation results of fixed parameters fall into the "high variance - low stability" trap, seriously weakening the efficiency and accuracy of the above traditional compartment models applied to pharmacokinetic parameter analysis.

[0030] Based on the above, the actual contradictions between the simplified assumptions of traditional compartment models and physiological reality, as well as the low utilization efficiency of PET-CT scanning equipment, have all led to bottlenecks in complex pharmacokinetic analyses. There is an urgent need to introduce new modeling methods to improve the ability to depict real biological movement processes. Therefore, an embodiment of this application discloses a dynamic PET-CT pharmacokinetic fusion modeling method that combines mechanism and data.

[0031] Referring to Figure 1 , a dynamic PET-CT pharmacokinetic fusion modeling method that combines mechanism and data includes: S100, obtaining all the original image data in historical cases and constructing a training dataset based on all the original image data; Among them, historical cases refer to the previous PET-CT examination records of subjects, usually including the complete dynamic PET-CT scan data from the start to the end of the injection of radioactive tracers (covering the early rapid distribution stage and the later metabolic clearance stage), the basic information of the subjects (such as age, weight, medical history), radioactive dose, scan parameters (such as time resolution), and region of interest (ROI) annotations (such as tumor lesions, organ boundaries); the original image data refers to the four-dimensional dynamic image signals containing the time dimension collected by the PET-CT device, and the original image data is used to reflect the distribution intensity of radioactive tracers in the subject's body at different time points. Specifically, constructing a training dataset based on all the original image data in S100 includes the following steps: S110, performing spatial alignment on the original image data between each dynamic frame to obtain registered image data after spatial alignment; Among them, the original image data in this embodiment is represented by , and the registered image data is represented by . Specifically, the relationship between the original image data and the registered image data can be expressed as follows: ; where T is the spatial transformation matrix calculated by the inter-frame registration algorithm (such as rigid transformation, affine transformation, or non-rigid deformation model), which is used to correct the displacement or deformation errors caused by factors such as breathing and body movement during the dynamic PET-CT scan of the subject; are the spatial coordinates and time coordinates, t corresponds to each time frame in the dynamic scan, corresponds to the specific position of the internal tissue or organ, x and y usually correspond to the two-dimensional plane coordinates of the image, representing the rows and columns of the scan tomogram, used to locate the pixel position within a certain plane, and z corresponds to the depth coordinate of the image, representing the stacking order of different scan tomograms, used to locate the voxel position in three-dimensional space; / represents the change in the radioactive signal intensity at spatial at time t.

[0032] Through this registration process, the images at different time points are ensured to be aligned in the spatial dimension, so as to eliminate the interference of motion artifacts on the subsequent analysis of pharmacokinetic parameters, laying a foundation for accurately extracting the tissue time-activity curve.

[0033] S120. Obtain the plasma input function corresponding to each case according to the original image data between all dynamic frames in each historical case. Among them, considering that the heart is the core of blood circulation and the signal in the heart region can directly reflect the dynamic concentration of the radioactive tracer in the plasma, the target region is selected as the heart region, and the plasma input function is used to reflect the dynamic concentration of the radioactive tracer in the plasma of the subjects in different historical cases. Specifically, S120 mainly extracts the signal in the target region of the registered image data as the plasma input function, and the plasma input function is as follows: ; Among them, is the plasma input function; is the number of voxels corresponding to the heart region, which is mainly obtained by segmenting the heart region in the image and counting the total number of voxels; i is the voxel index of the region, corresponding to the spatial coordinates of each voxel in the heart region .

[0034] It should be noted that in medical image processing, the calculation of the number of voxels usually follows the following steps: For a specific region, including the heart region in step S120 and the region of interest in subsequent step S130, its division first requires segmenting the corresponding region from the registered image data. After spatial alignment, the same spatial region in each registered image data corresponds to the same tissue / organ region. Therefore, in this embodiment, common region segmentation methods include but are not limited to: manual delineation (manually circling the heart boundary on the image by an experienced physician), threshold segmentation (setting a threshold to extract the target region based on the difference in radioactive intensity between the heart tissue and the surrounding tissue), template matching (using a pre-defined heart template, such as an elliptical model, for fitting), deep learning segmentation (automatically identifying the heart region using networks such as U-Net), etc.; the segmentation result is usually represented as a binary mask : ; After obtaining the mask of the heart region, the number of voxels can be calculated by the following formula: , that is, counting all voxels with a value of 1 in the mask.

[0035] S130. Extract the original tissue time-activity curve within the preset region of interest. Among them, the region of interest refers to the target tissue area specified by the user according to the diagnosis or research needs, such as tumor lesions, organs such as the liver / kidney, etc., which is used to reflect the dynamic process of the concentration of the radioactive tracer changing with time in a specific region. The principle of the segmentation method for the region of interest is the same as that for the heart region in step S120, which is a prior art and will not be elaborated here. Specifically, the extraction formula for the original tissue time-activity curve is as follows: ; Among them, is the number of voxels corresponding to the region of interest, which is mainly obtained by segmenting the region of interest in the image and counting the total number of voxels; i is the voxel index of this region, corresponding to the spatial coordinates of each voxel in the region of interest .

[0036] S140. Normalize the original tissue time-activity curve to obtain the standard tissue time-activity curve; Among them, normalization is mainly used to improve the convergence of the model and cross-individual comparability. The standard tissue time-activity curve is used to reflect the concentration dynamics of the radioactive tracer in the region of interest of subjects with different historical cases. Specifically, the normalization formula for the standard tissue time-activity curve is as follows: ; Among them, is the standard tissue time-activity curve; D is the dose of the injected radioactive tracer, and W is the body weight of the subject.

[0037] S150. Divide the original image data of each historical case into two parts to obtain early image data and full-course image data; Among them, the original image data in step S100 is a continuous long-time dynamic sequence. The division in this embodiment into two parts mainly refers to intercepting different stage data of the same dynamic sequence (i.e., the original image data) according to the time range. The early image data is used to reflect the actual drug distribution dynamics in the initial stage after the subject is injected with the radioactive tracer; the full-course image data is used to reflect the actual drug distribution dynamics of the subject from the initial stage until the end stage after being injected with the radioactive tracer. In this embodiment, taking 0 - 10 minutes as the initial stage and 0 - 60 minutes as the end stage as an example, the standard tissue time-activity curve corresponding to the early image data can be denoted as ; the standard tissue time-activity curve corresponding to the full-course image data can be denoted as , where coincides exactly with at ; in this embodiment takes 10 minutes, takes 60 minutes.

[0038] The early imaging data in this embodiment is acquired with high temporal resolution. For example, within 0 - 10 minutes after the subject is injected with a radioactive tracer, one frame of image is acquired every 10 - 30 seconds to capture the process of the tracer rapidly spreading to tissues with the blood flow, such as the initial uptake of 18F - FDG by the myocardium; the whole - course imaging data covers the early distribution, mid - stage metabolic transformation, and late - stage clearance and excretion phases. During this complete pharmacokinetic movement process, the temporal resolution of the imaging data decreases with the scanning progress. For example, within the subsequent 10 - 60 minutes, the acquisition frequency is reduced to one frame of image every 5 - 10 minutes, which is used to characterize the metabolic and clearance processes of the tracer. The whole - course imaging data covers the early imaging data and is used to provide kinetic details at the metabolic end (such as the retention of the tracer by tumors or the clearance trend of the tracer by the liver and kidneys).

[0039] S160, constructing a training dataset based on the early imaging data, whole - course imaging data, plasma input function, and standard tissue time - activity curve of each historical case; Among them, the training dataset is expressed as follows: ; Among them, represents the training dataset; j represents the j - th case, and N represents the number of subjects, that is, the number of cases; represents the standard tissue time - activity curve corresponding to the early imaging data of the j - th case; represents the standard tissue time - activity curve corresponding to the whole - course imaging data of the j - th case; represents the plasma input function corresponding to the j - th case.

[0040] S200, constructing a dynamic pharmacokinetic model based on a mechanism model and a deep - learning network, and completing the training of the dynamic pharmacokinetic model according to the training dataset; Among them, the dynamic pharmacokinetic model serves as the physical basis of the entire fusion architecture and directly reflects the transport and metabolic processes of the tracer in the subject. The construction of the dynamic pharmacokinetic model uses a mechanism model with clear biophysical significance, such as a two - compartment model, as the physical framework, and combines the dynamic compensation ability of the deep - learning network to form a hybrid architecture of "mechanism constraint + deep learning"; specifically: Regarding using the mechanism model as the basic physical framework, in this embodiment, mainly taking the two - compartment model as an example, the differential equation of the dynamic pharmacokinetic model in this embodiment can be expressed as: ; Among them, represents the concentration of the unbound radioactive tracer in the tissue, that is, the concentration of the tracer freely diffusing in the extracellular fluid of the tissue cells; is the concentration of the tracer that binds or becomes a metabolite in the tissue, i.e., the concentration of the tracer that binds to tissue targets or enters the metabolic pathway; is the transport rate of the tracer into the tissue, is the transport rate of the tracer out of the tissue, is the phosphorylation rate of the tracer, is the dephosphorylation rate of the tracer; is the plasma input function.

[0041] Since the output of the differential equation of the mechanism model is the theoretical concentration at continuous time, while the PET image data is the radioactive count signal at discrete time points, and the actual measurement includes complex factors such as the residual tracer in blood vessels and tissue-specific uptake differences. Therefore, in order to connect the output of the dynamic pharmacokinetic model in this embodiment with the actual PET image data (i.e., early image data and whole-course image data), that is, to ensure data consistency, this embodiment mainly constructs an observation function by introducing the vascular fraction to separate the tracer signals in blood vessels and tissues, so as to ensure that the final predicted output of the mechanism model is consistent with the actual PET signal (including plasma residue and tissue uptake); specifically, the observation function of this embodiment is expressed as follows: ; wherein, represents the predicted observed concentration of the mechanism model, that is, the radioactive activity concentration observed in the image data predicted by the mechanism model, with the unit of Bq / mL; represents the vascular fraction, reflecting the volume proportion of blood vessels in the tissue, which is a dimensionless parameter, and the value range is , represents that the tissue has no blood vessels at all (only contains extracellular fluid and cells), represents that the tissue is all blood vessels (an extreme case, usually does not exist); represents the concentration contribution of the unabsorbed plasma tracer in the tissue blood vessels to the PET signal, and the unit is the same as that of , and essentially it is a part of the concentration, used to reflect the weight of the tracer in blood vessels in the PET signal; represents the volume proportion of the non-vascular part of the tissue (i.e., the effective volume of extracellular fluid that can absorb the tracer); t is the independent variable, representing time; is the model parameter vector, including and other pharmacokinetic parameters. When t changes, outputs the predicted observed concentration at different time points. When changes, outputs the predicted observed concentration under different parameter settings.

[0042] Thus, in this embodiment, through the vascular fraction As the core connection link between the "theoretical pharmacokinetic process" and the "measured imaging data", it realizes the physical mapping between the output of the mechanistic model (tissue and plasma concentrations) and the actual PET signal. This design not only retains the biological interpretability of the two-compartment model but also enables flexible adjustment of the signal contributions from blood vessels and tissues, laying the foundation for the "mechanism-data fusion" modeling in this embodiment.

[0043] It should be noted that in the above observation function, the estimation of pharmacokinetic parameters can be transformed into a non-linear least squares fitting problem, and the optimal pharmacokinetic parameters are solved by minimizing the error between the predicted observed concentration and the actual PET observation data ; thus, in this embodiment, by continuously adjusting the parameters , the predicted observed concentration can be made as close as possible to the actual PET observation data at all time points t. This optimization effectively ensures that the dynamic pharmacokinetic model of this embodiment can fit the actual dynamic imaging data while having good biological interpretability and can be used for clinical tasks such as drug uptake analysis, metabolic flux estimation, and efficacy prediction of tumor lesions.

[0044] In addition, since traditional mechanistic models (such as the two-compartment model) adopt fixed parameters and simplified assumptions and cannot capture the time-variability of individual physiological states (such as fluctuations in metabolic enzyme activity) and complex physiological mechanisms (such as non-linear metabolic pathways, spatial heterogeneity noise), therefore, in order to improve the adaptability of the model to individual differences and the dynamic prediction accuracy, this embodiment mainly realizes the dynamic compensation for the mechanistic model through a dual neural network architecture composed of a parameter adjustment network and a residual prediction network. The specific steps include: S210, design a parameter adjustment network; among them, the parameter adjustment network is used to break through the limitation of the traditional model of "fixed parameters", that is, the limitation that it cannot accurately capture the time-variability of individual physiological states. In this embodiment, time-varying parameters are introduced to more accurately reflect the actual physiological process. Specifically, The calculation formula of is as follows: Among them, represents the basic parameter value at the population level, representing the pharmacokinetic parameters under the "average physiological state", such as the average transport rate of healthy people, which can be estimated from all the original imaging data of historical cases in S100; represents the model parameter vector corresponding to t; For capturing individual-specific and time-varying characteristics to adjust the value of the model parameter vector corresponding to t, is a time-varying parameter, which is a four-dimensional vector, that is, each element corresponds to an adjustment factor of a basic parameter; specifically: ; Among them, respectively correspond to the four core rate constants in the two-compartment model; if , it means that the tracer uptake rate of this subject is higher than the population average, and if , it means that the binding rate of the tracer in this subject decreases with time (such as receptor saturation).

[0045] Parameter adjustment network uses the tissue time-activity curve in the first 10 minutes before PET-CT and the plasma input function as inputs to predict the time-varying parameter ; Among them, are the parameter adjustment network parameters; represents the standard tissue time-activity curve between represents the plasma input function between is the time-varying parameter between , is the number of time points of

[0046] It should be noted that, in this embodiment, the parameter adjustment network integrates a convolutional layer, a bidirectional GRU, and an attention mechanism, can extract temporal features from a period of early PET data, maps them to four pharmacokinetic parameter adjustment factors on the complete time series through a multi-layer perceptron, and applies parameter constraints to ensure physiological rationality, realizing individualized time-varying parameter prediction to enhance the flexibility of the mechanistic model.

[0047] Specifically, the design in the parameter adjustment network is as follows: In this embodiment, the convolutional layer uses two 1D convolutions (kernel size = 3, number of channels = 32 → 64), equipped with BatchNorm and ReLU activation, to extract local features in the time series, such as the slope of the early drug uptake peak, the fluctuation frequency, etc., and convert the original signal into a high-dimensional feature vector; the hidden layer dimension of the bidirectional GRU is 128, which is used to capture the long-term dependencies of the time series data, such as the impact of the trend within the first 10 minutes on subsequent metabolism, and the bidirectional mechanism of the bidirectional GRU takes into account the information of both past and future time points, that is, although the input is only for a previous period of time (10 minutes), the GRU can also infer the subsequent trend (the next 50 minutes) through sequence modeling; an 8-head attention mechanism is used to calculate the attention weights between time points, for focusing on key time points (such as the peak time after tracer injection, the distribution equilibrium time), and suppressing the interference of noise on parameter prediction; In the application of the multi-layer perceptron, a structure of Flatten layer → fully connected layer (256 neurons, ReLU activation) → Dropout (0.2) is mainly adopted to map the features extracted by the convolution-GRU to the parameter adjustment space, generating a four-dimensional adjustment factor corresponding to the full-time course time points Regarding parameter constraints, in this embodiment, the correction term is mainly scaled to a reasonable range through α and β set based on physiological priors, such as etc., and it is ensured that is non-negative, , (if the metabolism is irreversible, then ).

[0048] In summary, the architecture design in the parameter adjustment network refers to the following during actual operation: Input layer: - # Tissue time-activity curve for the first 10 minutes - # Plasma input function for the first 10 minutes Feature extraction module: 1. Temporal feature encoding layer: - 1D convolutional layer (kernel size = 3, stride = 1, number of channels = 32) - BatchNormalization - ReLU activation - 1D convolutional layer (kernel size = 3, stride = 1, number of channels = 64) - BatchNormalization - ReLU activation 2. Temporal dynamic capture module: - Bidirectional GRU (hidden layer dimension = 128) - Self-attention mechanism (8-head attention), focusing on key time points 3. Feature integration layer: - Flatten - Dropout(0.2) - Dense(256, activation='relu') - Dropout(0.2) Parameter generation module: 4. Time-varying parameter prediction: -

[0049] - # Predict the changes of 4 parameters on the complete time series 5. Parameter constraint layer: - Tanh activation function (scaled to [-1, 1]) - Linear transformation (scaled to a reasonable range): ; where , is the scaling parameter set based on physiological prior knowledge Output layer: - # Four time-varying parameter adjustment factors.

[0050] S220, design a residual prediction network; Among them, the residual prediction network is used to break through the limitations of the "simplifying assumptions" of traditional models, that is, the limitations that cannot fully reflect complex physiological processes such as nonlinear metabolism, spatial heterogeneity, measurement noise, and unmodeled paths. In this embodiment, a residual correction term is introduced to compensate for the systematic bias of the mechanism model. The residual prediction network has the same input data as the parameter adjustment network in S210 and predicts the residual correction term covering the whole process The calculation formula of the residual correction term is as follows: ; where are the parameters of the residual prediction network; is the residual correction term, with a dimension of , is the number of time points.

[0051] It should be specifically noted that, in this embodiment, the residual prediction network combines multi-scale convolution, UNet, and bidirectional LSTM architectures, and can capture both local details and global trends from PET data over a previous period of time. Through an autoregressive mechanism, short sequences are extrapolated to a complete time series, and a smoothing constraint is applied to generate continuous residual correction terms, thereby effectively compensating for complex dynamic processes that the mechanism model fails to describe.

[0052] Specifically, the multi-scale convolution in this embodiment mainly uses parallel 1D convolution branches. Three branches respectively use 1D convolutions with kernel sizes of 3, 5, and 7 to capture local features at different time scales: small kernels of size 3 are used to extract short-term fluctuations, such as high-frequency changes in the early rapid distribution stage, and large kernels of size 5 or 7 are used to extract long-term trends, such as low-frequency features of late metabolic clearance; finally, the outputs of the three branches are concatenated to form a feature map containing multi-scale information; in this way, this embodiment can avoid the limitations of a single kernel size and enhance the network's ability to capture complex dynamic patterns; The adopted UNet structure is as follows: The encoding path is 3 layers of downsampling Conv1D, with the number of channels being 64 → 128 → 256 in sequence, stride = 2. Each downsampling expands the receptive field to extract global context features (such as the trend of the entire early time series); the decoding path is 3 layers of upsampling Conv1D, with the number of channels being 256 → 128 → 64 in sequence. The time dimension resolution is restored through transposed convolution and fused with the skip connection of the encoding path to integrate shallow details and deep semantics; in this way, this embodiment can achieve "compression - decompression" of time series data through the symmetric architecture of UNet, while extracting global trends, retaining local details such as the exact position of early peaks, etc.; the hidden layer dimension of the bidirectional LSTM = 128, which is used to process the fused feature sequence and capture bidirectional time series dependencies, such as the influence of future time points on the current prediction; the autoregressive prediction mechanism mainly uses Teacher-Forcing during the training stage, taking the true residual as the input for subsequent predictions to alleviate gradient disappearance, and recursively generating subsequent sequences based on the predicted residual terms during the inference stage to achieve extrapolation from to ; among them, during the initialization stage, the hidden state of the LSTM is initialized by the feature representation between to ensure the coherence of the extrapolation process with the known data.

[0053] In the application of the smoothing constraint, this embodiment applies a 1D convolution with a kernel size of 5 to smooth the predicted residuals, suppress high-frequency noise, and impose a stronger smoothing constraint on the residuals far from the known data points between assuming that the physiological process changes gradually over time to avoid sudden predictions.

[0054] In summary, the architecture design in the residual prediction network is referred to as follows during actual operation: Input layer: - # Tissue time-activity curve for the first 10 minutes - # Plasma input function for the first 10 minutes Feature extraction module: 1. Multi-scale feature extraction: - Three parallel 1D convolutional branches (kernel sizes are 3, 5, 7 respectively) - Each branch: Conv1D → BatchNorm → ReLU → MaxPooling - Feature connection layer: Concatenate 2. Context encoder: - 1D-UNet structure: - Encoding path: 3-layer downsampling Conv1D (number of channels increasing: 64 → 128 → 256) - Decoding path: 3-layer upsampling Conv1D (number of channels decreasing: 256 → 128 → 64) - Skip connections: between the corresponding encoding and decoding layers 3. Global-local feature fusion: - Global feature: GlobalAveragePooling → Dense(128) - Local feature: UNet output - Fusion layer: Concatenate → Conv1D Temporal prediction module: 4. Time extrapolation layer: - Bidirectional LSTM (128 units) - Autoregressive prediction mechanism: - Using Teacher-Forcing training - Initializing the hidden state using the feature representation of the first 10 minutes - Predictor network: Dense(1) → linear activation 5. Smoothing constraint layer: - Temporal smoothing filter: Conv1D (kernel size = 5, stride = 1) - Smoothing weight decay mechanism: The smoothing constraint is stronger at positions farther from the known data points Output layer: - # Complete residual term for 60 minutes

[0055] After completing the architecture design of the parameter adjustment network in step S210 and the residual prediction network in step S220 above, these two deep learning modules need to be integrated end-to-end with the mechanism model to form a complete dynamic pharmacokinetic model. At this time, the model architecture already has the dual-driving ability of "mechanism skeleton + data compensation", but the parameters of each component need to be optimized through a staged training process to ensure that: The basic parameters of the mechanism model have biophysical rationality; The parameters of the deep learning network can accurately capture the dynamic characteristics not covered by the mechanism model; When the two work together, they satisfy the balance between physiological constraints and prediction accuracy; Therefore, S200 of this embodiment further includes the following steps: S230, fusing and training the mechanism model and the deep learning network to obtain a dynamic pharmacokinetic model; Among them, in order to ensure that the fusion training starts from a meaningful physical solution space, it is necessary to provide biophysically reasonable initial values for the subsequent neural network. Therefore, S230 includes: S231, determining the population-based basic parameters of the mechanism model according to the standard tissue time-activity curve corresponding to the whole-course imaging data to complete pre-training; Among them, the population-based basic parameters are used to establish the initial kinetic framework of physiological laws. Specifically, the calculation formula of the population-based basic parameters is as follows: ; Among them, is the population-based basic parameter, corresponding to in S210; N represents the number of subjects; represents the basic parameter value at the population level; i represents the i-th case; represents the complete time series of the i-th case of the time point set; represents the tissue time-activity curve to be verified output by the mechanism model of the i-th case, which is obtained by solving the differential equation of the two-compartment model and depends on the parameter ; represents the standard tissue time-activity curve corresponding to the whole-course imaging data in the i-th case.

[0056] It should be noted that in step S231, this embodiment mainly determines a set of population-based basic parameters through non-linear least squares fitting, so that the tissue time-activity curve to be verified output by the mechanism model and the actually measured corresponding standard tissue time-activity curve have the smallest error to ensure that the model conforms to physiological laws such as mass conservation and transport direction.

[0057] Pre-trained Although it has met the physiological rationality at the population level, to prevent the neural network from violating the basic physical constraints during optimization, this embodiment mainly completes the training by sequentially training the parameter adjustment network and the residual prediction network in stages: S232. Fix the population-based parameters and train the parameter adjustment network according to the training dataset to obtain the time-varying parameter correction term; Specifically, the calculation formula of the time-varying parameter correction term is as follows: ; Wherein, represents the set of optimized optimal weight parameters of the parameter adjustment network, including neural network parameters such as convolutional layers, GRU layers, and attention mechanisms; represents the time-varying parameter correction term of the i-th case, which is a four-dimensional vector; represents the tissue time-activity curve to be verified output by the mechanism model after introducing the time-varying parameter, which is obtained by solving the differential equation of the two-compartment model; is a preset regularization hyperparameter used to control the amplitude of the time-varying parameter correction term to prevent from being over-adjusted to keep the corrected time-varying parameters within the physiologically reasonable range.

[0058] It should be noted that in step S232, if the tumor tissue of a certain subject is rich in blood vessels ( is relatively high), the parameter adjustment network may learn , indicating that its tracer uptake rate is higher than the population average, and thus dynamically adjusts the transport rate parameter of the tracer entering the tissue to adapt to individual differences.

[0059] After S232, the mechanism model and the parameter adjustment network have solved the problems of "fixed parameters" and "individual differences", while the residual network focuses on dealing with the limitations of the model structure itself (such as the enzyme saturation effect not modeled by the two-compartment model). Therefore, to avoid optimization confusion caused by multi-module coupling, S230 further includes: S233. Fix the population-based parameters and the trained parameter adjustment network, and train the parameter adjustment network according to the training dataset to obtain the residual correction term; Specifically, the calculation formula of the residual correction term is as follows: ; Wherein, represents the optimized optimal network weight parameters of the residual prediction network; represents the residual correction term of the i-th case at time; is a preset residual smoothness regularization hyperparameter, which is used to promote the continuous and smooth change of the residual over time, conforming to the continuous change characteristics of the physiological process; is a preset residual amplitude regularization hyperparameter, which is used to control the absolute value of the residual correction term to ensure that the residual is a fine-tuning rather than a dominant factor for the mechanistic model; represents the difference between the residual of the i-th case at the next moment and the residual at the current moment, that is, the sum of the squared differences of the residuals at adjacent time points, which is used to penalize mutations.

[0060] It should be noted that in step S233, if the measured PET signal shows abnormal fluctuations in the later stage (possibly caused by PET device noise), the residual prediction network will learn the corresponding residual correction term, but the amplitude constraint will limit its overreaction, and the smoothing constraint will ensure that the fluctuations do not violate the physiological continuity.

[0061] S234, perform end-to-end fine-tuning on the entire architecture, and jointly optimize the mechanistic model, the parameter adjustment network, and the residual prediction network to obtain a trained dynamic pharmacokinetic model; Specifically, the calculation formula for joint optimization is as follows: ; where, is the total loss function, which is used to balance the prediction accuracy, physical constraints, and model complexity; The specific calculation formula for the total loss function is as follows: ; where, is the adjustable weight factor for the physical constraint loss, which is used to adjust the strength of the physiological rationality constraint; is the adjustable weight factor for the regularization loss, which is used to control the model complexity; is the prediction loss, which is used to measure the fitting error between the model prediction value and the measured data; is the physical constraint loss, which is used to ensure that the model parameters conform to the physiological laws; is the regularization loss, which is used to prevent overfitting and enhance the model generalization ability; Specifically, the prediction loss, physical constraint loss, and regularization loss are respectively expressed as: ; where, M is the upper limit of the binding site occupancy ratio, that is, the upper limit of the rate ratio, usually M≥1; is the dynamic pharmacokinetic model, that is, the fusion model, and outputs the tissue time-activity curve to be verified; represents the ratio of the phosphorylation rate to the dephosphorylation rate, reflecting the metabolic reversibility; is the adjustable weight factor for the rate ratio constraint; For forcing the parameter to be non - negative, if , then the sequential loss is , otherwise it is 0.

[0062] S300, collect the actual image data of the current subject in the early period; Among them, the early period refers to the start stage after the subject is injected with the radioactive tracer. In this embodiment, the first 10 minutes is taken as an example; the actual image data is the four - dimensional dynamic image signal containing the time dimension collected by the current PET - CT device, which is essentially equivalent to the early image data in S100, and its acquisition method will not be elaborated here.

[0063] S400, input the actual image data into the trained dynamic pharmacokinetic model to obtain the prediction result.

[0064] Among them, the prediction result is a predicted tissue time - activity curve, and the predicted tissue time - activity curve is used to reflect the dynamic drug concentration obtained by prediction in the preset region of interest from the start stage to the end stage; referring to Figure 2 , when the actual image data is input into the trained dynamic pharmacokinetic model, the dynamic pharmacokinetic model can perform the following steps: ① Based on the same step principle as the above steps S100 - S200, extract the standard tissue time - activity curve of the region of interest in the actual image data and the plasma input function ; Among them, the standard tissue time - activity curve is used to reflect the early tracer distribution dynamics in the region of interest (i.e., a certain tissue / organ), such as the uptake rate, peak time, etc.; the plasma input function is used to reflect the hemodynamic characteristics of different subject individuals through the change of tracer concentration in the blood of the heart region, such as cardiac output, tracer injection rate, etc. The plasma input function can assist the dynamic pharmacokinetic model to capture the influence of inter - individual blood flow differences on pharmacokinetics; both the standard tissue time - activity curve and the plasma input function are used as the core input features of the subsequent parameter adjustment network and the residual prediction network; and in this embodiment, if it is difficult to accurately obtain the plasma input function of an individual, an input function model based on population statistics can also be used for estimation; ② Through the parameter adjustment network complete the following steps: ② / 1 According to the trained network parameters in step S232, the standard tissue time - activity curve, and the plasma input function, predict the time - varying parameter correction term: .

[0065] ② / 2 Update the population-based parameters of the mechanism model according to the time-varying parameter correction term to obtain the basic parameters corresponding to the current subject: 。

[0066] ② / 3 Solve the mechanism model according to the time-varying parameter and the basic parameters to obtain the time-activity curve of the tissue to be verified based on the actual image data of the subject in the early stage: , where represents the solution process of the ordinary differential equation.

[0067] ③ Through the residual prediction network Complete the following steps: ③ / 1 According to the trained network parameters in step S233 、the standard tissue time-activity curve and the plasma input function, predict the residual correction term: 。

[0068] ③ / 2 Superimpose the time-activity curve of the tissue to be verified output by the mechanism model and the residual correction term to obtain the fused prediction result, that is, the predicted tissue time-activity curve: 。

[0069] Based on the same inventive concept as above, an embodiment of the present application also discloses an intelligent terminal, which includes a memory and a processor. At least one instruction, at least one program, a code set or an instruction set is stored in the memory, and the at least one instruction, at least one program, a code set or an instruction set is loaded and executed by the processor to implement the dynamic PET-CT pharmacokinetic fusion modeling method combining mechanism and data provided in the above method embodiment.

[0070] Based on the same inventive concept as above, an embodiment of the present application also discloses a computer-readable storage medium, in which at least one instruction, at least one program, a code set or an instruction set is stored, and the at least one instruction, at least one program, a code set or an instruction set can be loaded and executed by a processor to implement the dynamic PET-CT pharmacokinetic fusion modeling method combining mechanism and data provided in the above method embodiment.

[0071] In addition, in order to ensure the accuracy of the dynamic pharmacokinetic model as much as possible, this embodiment can also periodically verify the model through evaluation and verification. Among them, for quantitative evaluation, MAE and R2 of curve prediction accuracy can be selected, for physiological rationality test, non-negative parameter constraint and rate ratio constraint can be selected, and for qualitative evaluation, sensitivity, specificity and AUC can be used; the above evaluation means are prior arts and will not be elaborated here.

[0072] Refer toFigure 3 , this application also discloses an electronic device. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502; Among them, the communication bus 502 is used to realize the connection and communication between these components.

[0073] Among them, the user interface 503 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 503 may also include a standard wired interface and a wireless interface.

[0074] Among them, the network interface 504 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0075] Among them, the processor 501 may include one or more processing cores. The processor 501 connects various parts within the entire server through various interfaces and lines. By running or executing instructions, programs, code sets, or instruction sets stored in the memory 505, and by calling data stored in the memory 505, it executes various functions of the server and processes data. Optionally, the processor 501 may be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), or programmable logic array (PLA). The processor 501 may integrate one or several combinations of a central processing unit (CPU), a graphics processing unit (GPU), and a modem, etc. Among them, the CPU mainly processes the operating system, user interface, and application programs, etc.; the GPU is responsible for the rendering and drawing of the content to be displayed on the display screen; the modem is used to process wireless communication. It can be understood that the above modem may not be integrated into the processor 501 and may be implemented separately through a single chip.

[0076] Among them, the memory 505 may include a Random Access Memory (RAM), or may also include a Read-Only Memory. Optionally, the memory 505 includes a non-transitory computer-readable storage medium. The memory 505 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 505 may include a program storage area and a data storage area. Among them, the program storage area may store instructions for implementing the operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-mentioned method embodiments, etc.; the data storage area may store the data involved in the above-mentioned method embodiments. Optionally, the memory 505 may also be at least one storage device located far from the aforementioned processor 501.

[0077] In the memory 505 as a computer storage medium, there may be included an operating system, a network communication module, a user interface module, and an application program for a dynamic PET-CT pharmacokinetic fusion modeling method that combines mechanism and data.

[0078] In Figure 3 In the electronic device 500 shown, the user interface 503 is mainly used to provide an input interface for the user and obtain the data input by the user; while the processor 501 can be used to call the application program for a dynamic PET-CT pharmacokinetic fusion modeling method that combines mechanism and data stored in the memory 505. When executed by one or more processors 501, the electronic device 500 is caused to execute the method as described in one or more of the above embodiments. It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that this application is not limited by the described action sequence, because according to this application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0079] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0080] In several embodiments provided by the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some service interfaces. The indirect coupling or communication connection of the device or unit can be in an electrical or other form.

[0081] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place, or can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0082] In addition, the functional units in each embodiment of the present application can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0083] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to enable a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the methods in each embodiment of the present application. And the aforementioned memory includes: various media such as USB flash drives, mobile hard disks, magnetic disks or optical discs that can store program codes.

[0084] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure still fall within the scope covered by the present disclosure. Those skilled in the art will easily think of other implementation schemes of the present disclosure after considering the specification and the disclosure of the practical truth.

[0085] The present application aims to cover any variations, uses or adaptive changes of the present disclosure. These variations, uses or adaptive changes follow the general principles of the present disclosure and include common general knowledge or conventional technical means in the technical field not recorded in the present disclosure. The specification and the embodiments are only regarded as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A dynamic PET-CT pharmacokinetic fusion modeling method combining mechanism and data, characterized in that, Including: Obtain all the original image data in historical cases, and construct a training dataset based on all the original image data; Construct a dynamic pharmacokinetic model based on a mechanism model and a deep learning network, and complete the training of the dynamic pharmacokinetic model according to the training dataset; Collect the actual image data of the current subject in the early stage; Input the actual image data into the trained dynamic pharmacokinetic model to obtain a prediction result, where the prediction result is a predicted tissue time-activity curve; The constructing the training dataset based on all the original image data includes: According to the original image data between all dynamic frames in each historical case, obtain a plasma input function and a standard tissue time-activity curve corresponding to each case; the plasma input function is used to reflect the concentration dynamics of the radioactive tracer in the plasma of the subjects in different historical cases, and the standard tissue time-activity curve is used to reflect the concentration dynamics of the radioactive tracer in the region of interest of the subjects in different historical cases; Divide the original image data of each historical case into two parts to obtain early image data and full-course image data; Construct a training dataset based on the early image data, full-course image data, plasma input function, and standard tissue time-activity curve of each historical case.

2. The dynamic PET-CT pharmacokinetic fusion modeling method according to claim 1, wherein The obtaining the plasma input function and the standard tissue time-activity curve corresponding to each case according to the original image data between all dynamic frames in each historical case includes: Perform spatial alignment on the original image data between each dynamic frame in each historical case to obtain registered image data after spatial alignment; Extract the signal in the target area of the registered image data as the plasma input function; Extract the original tissue time-activity curve within a preset region of interest; Normalize the original tissue time-activity curve to obtain a standard tissue time-activity curve.

3. The dynamic PET-CT pharmacokinetic fusion modeling method according to claim 1, wherein The dynamic PET-CT pharmacokinetic model includes a mechanism model and several deep learning networks, and the deep learning networks include a parameter adjustment network and a residual prediction network; the mechanism model is the basic physical framework of the dynamic PET-CT pharmacokinetic model; The parameter adjustment network is used to obtain a tissue time-activity curve to be verified according to the plasma input function and the standard tissue time-activity curve corresponding to the actual image data; The residual prediction network is used to obtain the predicted tissue time-activity curve according to the plasma input function and the standard tissue time-activity curve corresponding to the actual image data.

4. The dynamic PET-CT pharmacokinetic fusion modeling method according to claim 3, wherein The parameter adjustment network is specifically used to obtain time-varying parameters according to the plasma input function and the standard tissue time-activity curve corresponding to the actual image data, and correct the population-based parameters of the two-compartment differential equation in the mechanism model according to the time-varying parameters; the parameter adjustment network is further used to drive the mechanism model to solve the two-compartment differential equation to obtain a tissue time-activity curve to be verified; The residual prediction network is specifically configured to obtain a residual correction term based on the plasma input function corresponding to the actual imaging data and the standard tissue time-activity curve, and perform non-linear compensation on the tissue time-activity curve to be verified according to the residual correction term to obtain the predicted tissue time-activity curve.

5. The dynamic PET-CT pharmacokinetic fusion modeling method according to claim 1, wherein The early imaging data is used to reflect the actual drug distribution dynamics in the initial stage after the subject is injected with the radioactive tracer; the whole-course imaging data is used to reflect the actual drug distribution dynamics from the initial stage to the end stage after the subject is injected with the radioactive tracer; the predicted tissue time-activity curve is used to reflect the drug concentration dynamics obtained by prediction in the region of interest from the initial stage to the end stage.

6. The dynamic PET-CT pharmacokinetic fusion modeling method according to claim 5, characterized in that, Constructing a dynamic pharmacokinetic model based on a mechanism model and a deep learning network, and completing the training of the dynamic pharmacokinetic model according to the training data set includes: Designing a parameter adjustment network, where the parameter adjustment network is used to accurately reflect the actual physiological process; Designing a residual prediction network, where the residual prediction network is used to compensate for the systematic bias of the mechanism model; Based on the early imaging data, the whole-course imaging data, the plasma input function, and the standard tissue time-activity curve, jointly train the mechanism model, the parameter adjustment network, and the residual prediction network to obtain a dynamic pharmacokinetic model.

7. The dynamic PET-CT pharmacokinetic fusion modeling method according to claim 6, characterized in that Based on the early imaging data, the whole-course imaging data, the plasma input function, and the standard tissue time-activity curve, jointly training the mechanism model, the parameter adjustment network, and the residual prediction network to obtain a dynamic pharmacokinetic model includes: Determining the initial population-based parameters of the mechanism model according to the standard tissue time-activity curve corresponding to the whole-course imaging data to complete pre-training; Fixing the population-based parameters, and training the parameter adjustment network according to the standard tissue time-activity curve corresponding to the whole-course imaging data to obtain a time-varying parameter correction term; Fixing the population-based parameters and the trained parameter adjustment network, and training the parameter adjustment network according to the standard tissue time-activity curve corresponding to the whole-course imaging data to obtain a residual correction term; Performing end-to-end fine-tuning on the entire architecture, and jointly optimizing the mechanism model, the parameter adjustment network, and the residual prediction network to obtain a trained dynamic pharmacokinetic model.

8. The dynamic PET-CT pharmacokinetic fusion modeling method according to claim 7, wherein In the step of determining the population-based parameters of the mechanism model, a set of initial population-based parameters is determined by non-linear least squares fitting.

9. The dynamic PET-CT pharmacokinetic fusion modeling method according to claim 1, wherein The dynamic pharmacokinetic model realizes the data unification of the predicted tissue time-activity curve and the predicted observed concentration output by the mechanism model through an observation function, and the observation function is as follows: ； Among them, represents the predicted observed concentration output by the mechanism model; represents the vascular fraction, which reflects the volume proportion of blood vessels in the tissue and is a dimensionless parameter, ; represents the concentration contribution of the unextracted plasma tracer in the tissue blood vessels to the PET signal, and is used to reflect the weight of the tracer in the blood vessels in the PET signal; represents the volume proportion of the non-vascular part of the tissue; is the independent variable and represents time; is the model parameter vector; represents the concentration of the unbound radioactive tracer in the tissue, that is, the freely diffusing tracer in the extracellular fluid of the tissue cells; represents the plasma input function.

10. A computer-readable storage medium, characterized in that, At least one instruction, at least one program, a code set or an instruction set is stored in the readable storage medium, and the at least one instruction, at least one program, the code set or the instruction set is loaded and executed by a processor to implement a dynamic PET-CT pharmacokinetic fusion modeling method combining mechanism and data as described in any one of claims 1 to 9.

Images