Small animal living body multi-modal imaging system and method

Abstract

The invention discloses a multi-modal imaging system and method for a small animal living body. The system comprises a multi-modal fusion imaging module which is responsible for multi-modal image acquisition and primary processing; the real-time dynamic monitoring module is used for capturing bioluminescence / fluorescence signals in real time through a photon counting detector, receiving optical, nuclear medicine and magnetic resonance data, correcting motion artifacts based on a multi-scale space-time registration engine and adjusting scanning parameters through a real-time pharmacokinetic-physiological feedback mechanism; the low-radiation and biological compatible module is used for automatically optimizing the dosage of a tracer agent according to the weight of the animal, the scanning part and historical data by using a dosage prediction model; and the multi-modal data fusion module is used for performing non-rigid registration of multi-modal images based on a Transform cross-modal registration network, constructing a multi-species pharmacokinetic knowledge graph by utilizing multi-species metabolism chip data, integrating mouse, dog, primate and humanized liver chip data, and mining a cross-species metabolism rule through a graph neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal imaging, and in particular to a small animal living multimodal imaging system and method. Background Art

[0002] Small animal in vivo imaging technology is a non-invasive imaging method that can observe biological processes and drug distribution in real time in living animals. With the continuous development of molecular biology, cell biology, and imaging technology, small animal in vivo multimodal imaging technology has emerged. This technology combines multiple imaging modalities, such as optical imaging, magnetic resonance imaging (MRI), positron emission tomography (PET), etc., and can obtain structural and functional information of organisms at different scales, providing a more comprehensive and accurate means for pharmacokinetic research. Small animal in vivo multimodal imaging technology can observe the distribution and metabolic processes of drugs in the body in real time. Through technologies such as fluorescence imaging and bioluminescence imaging, the distribution and accumulation of drugs in tumor tissues, organs, and other parts can be tracked, providing important data for pharmacokinetic research.

[0003] In multimodal imaging, images from different modalities (such as PET and MRI) exhibit significant geometric distortion and nonlinear mapping relationships due to differences in imaging principles. Furthermore, physiological activities such as tumor growth and organ peristalsis in live animal experiments can cause complex image deformations. Traditional registration methods struggle to handle these nonlinear changes, resulting in large registration errors. Therefore, a system and method for in vivo multimodal imaging of small animals are proposed. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings of the existing technology in single-task learning methods, that is, the rating information of different dimensions may not be considered or fully utilized separately, and thus may not accurately reflect the specific evaluations and preferences of audiences or film critics in different dimensions. A small animal in vivo multimodal imaging system and method are proposed to address the shortcomings of the existing technology in single-task learning methods, that is, the rating information of different dimensions may not be considered or fully utilized separately, and thus may not accurately reflect the specific evaluations and preferences of audiences or film critics in different dimensions.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A small animal in vivo multimodal imaging system, comprising: Multimodal fusion imaging module: responsible for multimodal image acquisition and preliminary processing, including optical imaging (fluorescence / bioluminescence), nuclear medicine imaging (PET / SPECT), magnetic resonance imaging (MRI), and multi-species metabolic chip imaging; Real-time dynamic monitoring module: This module uses a photon counting detector to capture bioluminescence / fluorescence signals in real time, combines a deep learning dynamic sparse sampling algorithm to optimize data acquisition efficiency, accepts optical, nuclear medicine, and magnetic resonance data, and corrects motion artifacts based on a multi-scale spatiotemporal registration engine. It also adjusts scanning parameters through real-time pharmacokinetic-physiological feedback mechanisms (such as heart rate and blood pressure monitoring). Low Radiation and Biocompatibility Module: Utilizes a dose prediction model to automatically optimize tracer dosage based on animal weight, scan site, and historical data. A closed-loop anesthesia control system monitors anesthesia depth in real time (e.g., EEG / EMG feedback), dynamically adjusts isoflurane concentration, and reduces anesthesia-related physiological stress. Multimodal data fusion module: This module performs non-rigid registration of multimodal images based on the Transformer cross-modal registration network, constructs a multi-species pharmacokinetic knowledge graph using multi-species metabolic microarray data, integrates mouse, dog, primate, and humanized liver microarray data, and uses graph neural networks (GNNs) to mine cross-species metabolic patterns. The pharmacokinetic analysis module accepts the registered images provided by the multimodal data fusion module, combines the multi-scale PBPK model with tissue microstructure data (such as vascular density and cell type), calculates the distribution and metabolic rate of drugs in heterogeneous tissues, obtains pharmacokinetic parameters, learns cross-species metabolic kinetics through the GNN-ODE algorithm, and cross-validates the reliability of the model using a multi-species in vitro-in vivo joint validation platform; Real-time imaging optimization module: Uses a photonic neural network chip (PNN) to achieve spectral unmixing and noise suppression, shortens nuclear medicine reconstruction time to sub-seconds based on pharmacokinetic parameters combined with the quantum computing OSEM algorithm, and builds a multimodal data lake.

[0006] The above technical solution further includes: Preferably, the real-time dynamic monitoring module captures bioluminescence / fluorescence signals in real time through a photon counting detector and optimizes data acquisition efficiency by combining a deep learning dynamic sparse sampling algorithm, comprising the following steps: Signal acquisition: Photon counting detectors record photon arrival times in time-resolved mode and location , forming the original photon event stream , where N is the total number of photons; Noise suppression: Filter out background noise using a Poisson noise model ,in, is the background photon rate, Δt is the time window; Sparse sampling strategy: predicting signal importance weight based on U-Net convolutional neural network , define the sparse sampling mask , where τ is the threshold (e.g. 0.7); Compressed sensing reconstruction: sparsely sampled data , L1 regularization is used to reconstruct , where A is the measurement matrix and λ is the regularization parameter; Dynamically adjust the sampling rate: Adaptively adjust the sparse rate r according to the signal change rate ΔS: ,in, is the adjustment coefficient (0.8); Physiological feedback integration: combining heart rate H(t) and respiratory rate R(t) to modify sampling strategy ,in, is the feedback intensity.

[0007] Preferably, the real-time dynamic monitoring module receives optical, nuclear medicine, and magnetic resonance data and corrects motion artifacts based on a multi-scale spatiotemporal registration engine, and adjusts scanning parameters through a real-time pharmacokinetic-physiological feedback mechanism (such as heart rate and blood pressure monitoring), including the following steps: Initial alignment: based on MRI anatomical images , through rigid body transformation , aligned PET / optical images ,in, is the rotation matrix, is the translation vector; Dynamic non-rigid registration: using a multi-scale B-spline free-form deformation (FFD) model , by minimizing the mutual information loss Optimize deformation parameters : ,in, is the entropy function; Motion artifact correction: combined with breathing / heartbeat signals , building a spatiotemporal motion model , to compensate for non-rigid deformation ; Physiological parameter monitoring: Real-time monitoring of heart rate through closed-loop anesthesia control system and blood pressure , the data sampling frequency is 100Hz; Pharmacokinetic model-driven parameter optimization: Calculating drug concentrations in tissues based on multiscale PBPK models :The calculation formula is expressed as ,in, is the blood perfusion rate parameter; Dynamic adjustment of scanning parameters: Adjust PET scanning time according to physiological parameters and pharmacokinetic status and tracer injection rate , the calculation formula is expressed as in, is the empirical coefficient, is the baseline blood pressure.

[0008] Preferably, the low radiation and biocompatible module automatically optimizes the tracer dosage based on the animal's weight, scan site, and historical data using a dose prediction model, wherein the dose prediction model is constructed using a random forest; The low-radiation and biocompatible module monitors the depth of anesthesia in real time (e.g., EEG / EMG feedback) through a closed-loop anesthesia control system, dynamically adjusts the isoflurane concentration, and reduces anesthesia-related physiological stress, including the following steps: Closed-loop anesthesia control system input parameters: EEG / EMG signal (real-time monitoring of anesthesia depth), EEG frequency band power, EMG amplitude; output: target anesthesia depth (T, categorical variable, such as light anesthesia, moderate anesthesia, deep anesthesia); Isoflurane concentration adjustment: Dynamic adjustment of isoflurane concentration based on PID control algorithm: in, is the initial concentration, is the error term, is the PID control parameter (needs to be calibrated through experiments).

[0009] Anesthesia depth score calculation: EEG / EMG signals are integrated using a weighted sum model.

[0010] Preferably, the multimodal data fusion module performs the following specific steps for non-rigid registration of multimodal images based on the Transformer cross-modal registration network: Input data preprocessing and feature extraction: Segment the multimodal image into fixed-size patches and extract feature vectors through convolutional layers; Transformer cross-modal registration network construction: The Transformer self-attention mechanism is used to capture the nonlinear mapping relationship between PET and MRI features. The registration formula is expressed as: ; Deformation field generation and non-rigid registration: The registration features are mapped to a deformation field (DF) through a fully connected layer and applied to the PET image for non-rigid registration. The deformation field generation formula is: ; The image after registration is ; Loss function and optimization: The multimodal structural similarity index (MS-SSIM) is used as the loss function to measure the anatomical consistency between the registered image and the MRI, which is expressed as: in, , is the mean, , is the variance, is the covariance, is a constant.

[0011] Preferably, the multimodal data fusion module uses multi-species metabolic chip data to construct a multi-species pharmacokinetic knowledge graph, integrates mouse, dog, primate and humanized liver chip data, and mines cross-species metabolic patterns through a graph neural network (GNN), including the following steps: Multi-species metabolic microarray data acquisition and preprocessing: Metabolic data are collected through multi-species metabolic microarrays (such as mouse, dog, primate, and humanized liver microarrays), and Z-score normalization is used to eliminate dimensional differences; Construction of a multi-species pharmacokinetic knowledge graph: defining nodes (such as drugs, metabolites, enzymes) and edges (such as metabolic reactions, similarities between species), and calculating the similarity of metabolic pathways between species: in is the k-th metabolic reaction rate of species i; Graph Neural Network (GNN) mining cross-species metabolic patterns: node feature update formula: in, is the feature of node v in layer l, N(v) is the neighbor node, is the normalization coefficient, Weight matrix; learn the conservation and differences of metabolic pathways between species through GNN.

[0012] Preferably, the pharmacokinetic analysis module receives the registered images provided by the multimodal data fusion module, combines the multiscale PBPK model with tissue microstructure data, calculates the distribution and metabolic rate of the drug in heterogeneous tissues, obtains pharmacokinetic parameters, learns cross-species metabolic kinetics through the GNN-ODE algorithm, and uses the multi-species in vitro-in vivo joint verification platform to cross-validate the reliability of the model. The specific steps are as follows: The multimodal data fusion module outputs registered images: The multimodal data fusion module outputs a registered image sequence, including anatomical structure (MRI) and metabolic distribution (PET); Multiscale PBPK model construction and parameter initialization: The multiscale PBPK model divides tissues into multiple compartments (such as blood vessels, interstitial cells, and intracellular compartments), each of which contains the following parameters: blood perfusion rate (Unit: mL / min / g); tissue volume (Unit: mL); Drug partition coefficient : Parameter initialization formula: in, is the drug concentration in compartment i, is the metabolic rate; Tissue microstructure data integration: vascular density is measured by MRI perfusion imaging, cell type distribution is measured by tissue section staining and AI segmentation, and parameter correction is adjusted according to microstructure data and , improve the accuracy of heterogeneous tissue parameters (error <8%); The GNN-ODE algorithm learns cross-species metabolic dynamics: It combines the PBPK model differential equation with the graph neural network (GNN) to learn cross-species metabolic laws. The differential equation form is: Where C(t) is the drug concentration vector, is the PBPK parameter, G is the cross-species metabolic map (nodes are metabolites / genes, edges are regulatory relationships); GNN embedding update: in, is the embedding of node v, N(v) is the neighboring node; Use the Runge-Kutta method to solve the differential equation and predict the change of drug concentration over time; Cross-validation of a multi-species in vitro-in vivo combined validation platform: drug metabolism rates are measured using a liver chip (e.g., LC-MS to detect metabolite concentrations), PET / MRI data are collected in mouse / primate models, and pharmacokinetic parameters are calculated.

[0013] Cross-validation: Input: in vitro chip data (metabolic rate) and in vivo animal data (pharmacokinetic parameters), calculate the loss function, including mean square error (MSE) and Pearson correlation coefficient ( ): , in vitro-in vivo prediction error <10%, >0.85; Cross-species pharmacokinetic parameter output: Pharmacokinetic parameters: Volume of distribution , clearance rate CL, half-life ; Cross-species prediction: mapping of pharmacokinetic parameters from mouse to primate to human.

[0014] Preferably, a small animal in vivo multimodal imaging system and a small animal in vivo multimodal imaging method corresponding to the small animal in vivo multimodal imaging system include: Animal model selection and preparation: Select a small animal model (mice, rats), anesthetize it (isoflurane inhalation anesthesia), and fix it to the imaging platform. The depth of anesthesia is monitored in real time using a closed-loop anesthesia control system. Tracer injection and dose optimization: The tracer dosage (18F-FDG) was calculated based on the dose prediction model and administered via tail vein injection. Synchronous multimodal image acquisition: Use high-sensitivity photon counting detectors to collect fluorescence / bioluminescence signals, initiate PET / SPECT scans, combine deep learning dynamic sparse sampling algorithms to optimize data acquisition, and execute MRI sequences (T1 / T2-weighted imaging) to obtain anatomical structural information; Real-time physiological parameter monitoring: record heart rate, blood pressure and other physiological parameters through real-time dynamic monitoring module; Multimodal image registration and fusion: The original multimodal image is input into the Transformer cross-modal registration network to generate a registered image, which is then combined with the multi-species pharmacokinetic knowledge graph to construct a cross-modal metabolic joint map; Pharmacokinetic parameter quantification: The registered images are input into a multi-scale PBPK model to extract pharmacokinetic parameters. The model fit is optimized using the GNN-ODE algorithm, and the results are verified using a multi-species in vitro-in vivo joint validation platform. Real-time imaging optimization: Utilizes quantum computing OSEM algorithms to optimize PET reconstruction, shorten reconstruction time, and store analysis results in a multimodal data lake to support subsequent retrieval and reuse; Mining cross-species metabolic patterns: Analyzing metabolic differences among mouse, dog, primate, and humanized liver chips through a multi-species pharmacokinetic knowledge graph, and utilizing a federated learning framework to achieve cross-center model collaborative training and improve the universality of metabolic patterns. Model validation and error analysis: Compare predicted pharmacokinetic parameter values with actual measured values, calculate the error (<8%), and evaluate model reliability using a multi-species in vitro-in vivo combined validation platform; Clinical translational application: Optimize dosage design in new drug development based on cross-species metabolic patterns.

[0015] The present invention has the following beneficial effects: This invention introduces a Transformer-based cross-modal registration network, which captures the nonlinear mapping relationships between multimodal images and reduces registration errors under complex deformations. This allows fused images of modalities such as PET / MRI to more accurately reflect the correspondence between anatomical structure and metabolic function. The Transformer network is used to process optical / PET time series data in real time, compensating for non-rigid deformations caused by breathing / heartbeat, ensuring the continuity of image registration during dynamic monitoring, reducing errors caused by motion artifacts, and improving data quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a system architecture diagram of a small animal in vivo multimodal imaging system proposed by the present invention; Figure 2 This is a flow chart of a multimodal imaging method for small animal in vivo proposed by the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, a small animal in vivo multimodal imaging system includes: Multimodal fusion imaging module: responsible for multimodal image acquisition and preliminary processing, including optical imaging (fluorescence / bioluminescence), nuclear medicine imaging (PET / SPECT), magnetic resonance imaging (MRI), and multi-species metabolic chip imaging; Real-time dynamic monitoring module: This module uses a photon counting detector to capture bioluminescence / fluorescence signals in real time, combines a deep learning dynamic sparse sampling algorithm to optimize data acquisition efficiency, accepts optical, nuclear medicine, and magnetic resonance data, and corrects motion artifacts based on a multi-scale spatiotemporal registration engine. It also adjusts scanning parameters through real-time pharmacokinetic-physiological feedback mechanisms (such as heart rate and blood pressure monitoring). Low Radiation and Biocompatibility Module: Utilizes a dose prediction model to automatically optimize tracer dosage based on animal weight, scan site, and historical data. A closed-loop anesthesia control system monitors anesthesia depth in real time (e.g., EEG / EMG feedback), dynamically adjusts isoflurane concentration, and reduces anesthesia-related physiological stress. Multimodal data fusion module: This module performs non-rigid registration of multimodal images based on the Transformer cross-modal registration network, constructs a multi-species pharmacokinetic knowledge graph using multi-species metabolic microarray data, integrates mouse, dog, primate, and humanized liver microarray data, and uses graph neural networks (GNNs) to mine cross-species metabolic patterns. The pharmacokinetic analysis module accepts the registered images provided by the multimodal data fusion module, combines the multi-scale PBPK model with tissue microstructure data (such as vascular density and cell type), calculates the distribution and metabolic rate of drugs in heterogeneous tissues, obtains pharmacokinetic parameters, learns cross-species metabolic kinetics through the GNN-ODE algorithm, and cross-validates the reliability of the model using a multi-species in vitro-in vivo joint validation platform; Real-time imaging optimization module: Uses a photonic neural network chip (PNN) to achieve spectral unmixing and noise suppression, shortens nuclear medicine reconstruction time to sub-seconds based on pharmacokinetic parameters combined with the quantum computing OSEM algorithm, and builds a multimodal data lake.

[0019] In one embodiment, the real-time dynamic monitoring module captures bioluminescence / fluorescence signals in real time using a photon counting detector and optimizes data acquisition efficiency by combining a deep learning dynamic sparse sampling algorithm, including the following steps: Signal acquisition: Photon counting detectors record photon arrival times in time-resolved mode and location , forming the original photon event stream , where N is the total number of photons; Noise suppression: Filter out background noise using a Poisson noise model ,in, is the background photon rate, Δt is the time window; Sparse sampling strategy: predicting signal importance weight based on U-Net convolutional neural network , define the sparse sampling mask , where τ is the threshold (e.g. 0.7); Compressed sensing reconstruction: sparsely sampled data , L1 regularization is used to reconstruct , where A is the measurement matrix and λ is the regularization parameter; Dynamically adjust the sampling rate: Adaptively adjust the sparse rate r according to the signal change rate ΔS: ,in, is the adjustment coefficient (0.8); Physiological feedback integration: combining heart rate H(t) and respiratory rate R(t) to modify sampling strategy ,in, is the feedback intensity.

[0020] In one embodiment, the real-time dynamic monitoring module receives optical, nuclear medicine, and magnetic resonance data and corrects motion artifacts based on a multi-scale spatiotemporal registration engine, and adjusts scanning parameters through a real-time pharmacokinetic-physiological feedback mechanism (e.g., heart rate and blood pressure monitoring), including the following steps: Initial alignment: based on MRI anatomical images , through rigid body transformation Aligning PET / optical images in, is the rotation matrix, is the translation vector; Dynamic non-rigid registration: using a multi-scale B-spline free-form deformation (FFD) model , by minimizing the mutual information loss Optimize deformation parameters : ,in is the entropy function; Motion artifact correction: combined with breathing / heartbeat signals , building a spatiotemporal motion model , to compensate for non-rigid deformation ; Physiological parameter monitoring: Real-time monitoring of heart rate through closed-loop anesthesia control system and blood pressure , the data sampling frequency is 100Hz; Pharmacokinetic model-driven parameter optimization: Calculating drug concentrations in tissues based on multiscale PBPK models :The calculation formula is expressed as in, is the blood perfusion rate parameter; Dynamic adjustment of scanning parameters: Adjust PET scanning time according to physiological parameters and pharmacokinetic status and tracer injection rate The calculation formula is expressed as in, is the empirical coefficient, is the baseline blood pressure.

[0021] In one embodiment, the low radiation and biocompatible module automatically optimizes the tracer dosage based on the animal's weight, scan site, and historical data using a dose prediction model constructed using a random forest; The low-radiation and biocompatible module monitors the depth of anesthesia in real time (e.g., EEG / EMG feedback) through a closed-loop anesthesia control system, dynamically adjusts the isoflurane concentration, and reduces anesthesia-related physiological stress, including the following steps: Closed-loop anesthesia control system input parameters: EEG / EMG signal (real-time monitoring of anesthesia depth), EEG frequency band power, EMG amplitude; output: target anesthesia depth (T, categorical variable, such as light anesthesia, moderate anesthesia, deep anesthesia); Isoflurane concentration adjustment: Dynamic adjustment of isoflurane concentration based on PID control algorithm: in, is the initial concentration, is the error term, is the PID control parameter (needs to be calibrated through experiments).

[0022] Anesthesia depth score calculation: EEG / EMG signals are integrated using a weighted sum model.

[0023] Preferably, the multimodal data fusion module performs the following specific steps for non-rigid registration of multimodal images based on the Transformer cross-modal registration network: Input data preprocessing and feature extraction: Segment the multimodal image into fixed-size patches and extract feature vectors through convolutional layers; Transformer cross-modal registration network construction: The Transformer self-attention mechanism is used to capture the nonlinear mapping relationship between PET and MRI features. The registration formula is expressed as: ; Deformation field generation and non-rigid registration: The registration features are mapped to a deformation field (DF) through a fully connected layer and applied to the PET image for non-rigid registration. The deformation field generation formula is: ; The image after registration is ; Loss function and optimization: The multimodal structural similarity index (MS-SSIM) is used as the loss function to measure the anatomical consistency between the registered image and the MRI, which is expressed as: in, , is the mean, , is the variance, is the covariance, is a constant.

[0024] In one embodiment, the multimodal data fusion module uses multi-species metabolic chip data to construct a multi-species pharmacokinetic knowledge graph, integrates mouse, dog, primate, and humanized liver chip data, and mines cross-species metabolic patterns through a graph neural network (GNN), including the following steps: Multi-species metabolic microarray data acquisition and preprocessing: Metabolic data are collected through multi-species metabolic microarrays (such as mouse, dog, primate, and humanized liver microarrays), and Z-score normalization is used to eliminate dimensional differences; Construction of a multi-species pharmacokinetic knowledge graph: defining nodes (such as drugs, metabolites, enzymes) and edges (such as metabolic reactions, similarities between species), and calculating the similarity of metabolic pathways between species: in is the k-th metabolic reaction rate of species i; Graph Neural Network (GNN) mining cross-species metabolic patterns: node feature update formula: in, is the feature of node v in layer l, N(v) is the neighbor node, is the normalization coefficient, Weight matrix; learn the conservation and differences of metabolic pathways between species through GNN.

[0025] In one embodiment, the pharmacokinetic analysis module receives the registered images provided by the multimodal data fusion module, combines the multiscale PBPK model with tissue microstructure data, calculates the distribution and metabolic rate of the drug in heterogeneous tissues, obtains pharmacokinetic parameters, learns cross-species metabolic kinetics through the GNN-ODE algorithm, and uses a multi-species in vitro-in vivo joint validation platform to cross-validate the reliability of the model. The specific steps are as follows: The multimodal data fusion module outputs registered images: The multimodal data fusion module outputs a registered image sequence, including anatomical structure (MRI) and metabolic distribution (PET); Multiscale PBPK model construction and parameter initialization: The multiscale PBPK model divides tissues into multiple compartments (such as blood vessels, interstitial cells, and intracellular compartments), each of which contains the following parameters: blood perfusion rate (Unit: mL / min / g); tissue volume (Unit: mL); Drug partition coefficient : Parameter initialization formula: in, is the drug concentration in compartment i, is the metabolic rate; Tissue microstructure data integration: vascular density is measured by MRI perfusion imaging, cell type distribution is measured by tissue section staining and AI segmentation, and parameter correction is adjusted according to microstructure data and , improve the accuracy of heterogeneous tissue parameters (error <8%); The GNN-ODE algorithm learns cross-species metabolic dynamics: It combines the PBPK model differential equation with the graph neural network (GNN) to learn cross-species metabolic laws. The differential equation form is: Where C(t) is the drug concentration vector, is the PBPK parameter, G is the cross-species metabolic map (nodes are metabolites / genes, edges are regulatory relationships); GNN embedding update: in, is the embedding of node v, N(v) is the neighboring node; Use the Runge-Kutta method to solve the differential equation and predict the change of drug concentration over time; Cross-validation of a multi-species in vitro-in vivo combined validation platform: drug metabolism rates are measured using a liver chip (e.g., LC-MS to detect metabolite concentrations), PET / MRI data are collected in mouse / primate models, and pharmacokinetic parameters are calculated.

[0026] Cross-validation: Input: in vitro chip data (metabolic rate) and in vivo animal data (pharmacokinetic parameters), calculate the loss function, including mean square error (MSE) and Pearson correlation coefficient ( ): , in vitro-in vivo prediction error <10%, >0.85; Cross-species pharmacokinetic parameter output: Pharmacokinetic parameters: Volume of distribution , clearance rate CL, half-life ; Cross-species prediction: mapping of pharmacokinetic parameters from mouse to primate to human.

[0027] like Figure 2 As shown, a small animal in vivo multimodal imaging system and a small animal in vivo multimodal imaging method corresponding to the small animal in vivo multimodal imaging system include: Animal model selection and preparation: Select a small animal model (mice, rats), anesthetize it (isoflurane inhalation anesthesia), and fix it to the imaging platform. The depth of anesthesia is monitored in real time using a closed-loop anesthesia control system. Tracer injection and dose optimization: The tracer dosage (18F-FDG) was calculated based on the dose prediction model and administered via tail vein injection. Synchronous multimodal image acquisition: Use high-sensitivity photon counting detectors to collect fluorescence / bioluminescence signals, initiate PET / SPECT scans, combine deep learning dynamic sparse sampling algorithms to optimize data acquisition, and execute MRI sequences (T1 / T2-weighted imaging) to obtain anatomical structural information; Real-time physiological parameter monitoring: record heart rate, blood pressure and other physiological parameters through real-time dynamic monitoring module; Multimodal image registration and fusion: The original multimodal image is input into the Transformer cross-modal registration network to generate a registered image, which is then combined with the multi-species pharmacokinetic knowledge graph to construct a cross-modal metabolic joint map; Pharmacokinetic parameter quantification: The registered images are input into a multi-scale PBPK model to extract pharmacokinetic parameters. The model fit is optimized using the GNN-ODE algorithm, and the results are verified using a multi-species in vitro-in vivo joint validation platform. Real-time imaging optimization: Utilizes quantum computing OSEM algorithms to optimize PET reconstruction, shorten reconstruction time, and store analysis results in a multimodal data lake to support subsequent retrieval and reuse; Mining cross-species metabolic patterns: Analyzing metabolic differences among mouse, dog, primate, and humanized liver chips through a multi-species pharmacokinetic knowledge graph, and utilizing a federated learning framework to achieve cross-center model collaborative training and improve the universality of metabolic patterns. Model validation and error analysis: Compare predicted pharmacokinetic parameter values with actual measured values, calculate the error (<8%), and evaluate model reliability using a multi-species in vitro-in vivo combined validation platform; Clinical translational application: Optimize dosage design in new drug development based on cross-species metabolic patterns.

[0028] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A small animal in vivo multimodal imaging system, characterized in that: include: Multimodal fusion imaging module: responsible for multimodal image acquisition and preliminary processing, including optical imaging, nuclear medicine imaging, magnetic resonance imaging, and multi-species metabolic chip imaging; Real-time dynamic monitoring module: This module uses a photon counting detector to capture bioluminescence / fluorescence signals in real time, combines a deep learning dynamic sparse sampling algorithm to optimize data acquisition efficiency, accepts optical, nuclear medicine, and magnetic resonance data, and corrects motion artifacts based on a multi-scale spatiotemporal registration engine. Furthermore, it adjusts scanning parameters through a real-time pharmacokinetic-physiological feedback mechanism. Low Radiation and Biocompatible Module: Utilizes a dose prediction model to automatically optimize tracer dosage based on animal weight, scan site, and historical data. A closed-loop anesthesia control system monitors anesthesia depth in real time and dynamically adjusts isoflurane concentration. Multimodal data fusion module: Based on the Transformer cross-modal registration network, non-rigid registration of multimodal images is performed. Multi-species metabolic chip data are used to construct a multi-species pharmacokinetic knowledge graph. Mouse, dog, primate, and humanized liver chip data are integrated, and cross-species metabolic patterns are mined through graph neural networks. Pharmacokinetic Analysis Module: This module receives the registered images provided by the Multimodal Data Fusion Module, combines the multi-scale PBPK model with tissue microstructure data, calculates the distribution and metabolic rate of drugs in heterogeneous tissues, obtains pharmacokinetic parameters, and learns cross-species metabolic kinetics through the GNN-ODE algorithm. Real-time imaging optimization module: uses a photonic neural network chip to achieve spectral unmixing and noise suppression, shortens the nuclear medicine reconstruction time based on pharmacokinetic parameters and combines the quantum computing OSEM algorithm to build a multimodal data lake.

2. A small animal living multimodal imaging system according to claim 1, characterized in that: The real-time dynamic monitoring module uses a photon counting detector to capture bioluminescence / fluorescence signals in real time and combines a deep learning dynamic sparse sampling algorithm to optimize data acquisition efficiency, including the following steps: Signal acquisition: Photon counting detectors record photon arrival times in time-resolved mode and location , forming the original photon event stream , where N is the total number of photons; Noise suppression: Filter out background noise using a Poisson noise model ,in, is the background photon rate, Δt is the time window; Sparse sampling strategy: predicting signal importance weight based on U-Net convolutional neural network , define the sparse sampling mask , where τ is the threshold; Compressed sensing reconstruction: sparsely sampled data , L1 regularization is used to reconstruct , where A is the measurement matrix and λ is the regularization parameter; Dynamically adjust the sampling rate: Adaptively adjust the sparse rate r according to the signal change rate ΔS: ,in, is the adjustment coefficient; Physiological feedback integration: combining heart rate H(t) and respiratory rate R(t) to modify sampling strategy ,in, is the feedback intensity.

3. The small animal in vivo multimodal imaging system according to claim 1, characterized in that: The real-time dynamic monitoring module receives optical, nuclear medicine, and magnetic resonance data and corrects motion artifacts based on a multi-scale spatiotemporal registration engine, and adjusts scanning parameters through a real-time pharmacokinetic-physiological feedback mechanism, including the following steps: Initial alignment: based on MRI anatomical images , through rigid body transformation , aligned PET / optical images ,in, is the rotation matrix, is the translation vector; Dynamic non-rigid registration: using a multi-scale B-spline free-form deformation (FFD) model , by minimizing the mutual information loss Optimize deformation parameters : ,in, is the entropy function; Motion artifact correction: combined with breathing / heartbeat signals , building a spatiotemporal motion model , to compensate for non-rigid deformation ; Physiological parameter monitoring: Real-time monitoring of heart rate through closed-loop anesthesia control system and blood pressure , the data sampling frequency is 100Hz; Pharmacokinetic model-driven parameter optimization: Calculating drug concentrations in tissues based on multiscale PBPK models :The calculation formula is expressed as ,in, is the blood perfusion rate parameter; Dynamic adjustment of scanning parameters: Adjust PET scanning time according to physiological parameters and pharmacokinetic status and tracer injection rate , the calculation formula is expressed as in, is the empirical coefficient, is the baseline blood pressure.

4. The multimodal imaging system for small animals according to claim 1, characterized in that: The low radiation and biocompatible module automatically optimizes the tracer dosage based on the animal's weight, scan site, and historical data using a dose prediction model constructed using random forests; The low-radiation and biocompatible module monitors the depth of anesthesia in real time through a closed-loop anesthesia control system and dynamically adjusts the isoflurane concentration, including the following steps: Closed-loop anesthesia control system input parameters: EEG / EMG signal, EEG frequency band power, EMG amplitude; output: target anesthesia depth; Isoflurane concentration adjustment: Dynamic adjustment of isoflurane concentration based on PID control algorithm: in, is the initial concentration, is the error term, is the PID control parameter.

5. Calculation of anesthesia depth score: EEG / EMG signals are synthesized using a weighted sum model; The small animal in vivo multimodal imaging system according to claim 1, characterized in that: The multimodal data fusion module performs non-rigid registration of multimodal images based on the Transformer cross-modal registration network. Input data preprocessing and feature extraction: Segment the multimodal image into fixed-size patches and extract feature vectors through convolutional layers; Transformer cross-modal registration network construction: The Transformer self-attention mechanism is used to capture the nonlinear mapping relationship between PET and MRI features. The registration formula is expressed as: ； Deformation field generation and non-rigid registration: The registration features are mapped into a deformation field through a fully connected layer and applied to PET images for non-rigid registration. The deformation field generation formula is: ； The image after registration is ; Loss function and optimization: The multimodal structural similarity index (MS-SSIM) is used as the loss function to measure the anatomical consistency between the registered image and the MRI, which is expressed as: in, , is the mean, , is the variance, is the covariance, is a constant.

6. The small animal in vivo multimodal imaging system according to claim 1, characterized in that: The multimodal data fusion module uses multi-species metabolic chip data to construct a multi-species pharmacokinetic knowledge graph, integrates mouse, dog, primate, and humanized liver chip data, and mines cross-species metabolic patterns through graph neural networks. The module includes the following steps: Multi-species metabolic array data acquisition and preprocessing: Metabolic data were collected through multi-species metabolic arrays, and Z-score normalization was used to eliminate dimensional differences; Construction of a multi-species pharmacokinetic knowledge graph: defining nodes and edges, and calculating the similarity of metabolic pathways between species: in is the k-th metabolic reaction rate of species i; Graph neural network mining cross-species metabolic rules: node feature update formula: in, is the feature of node v in layer l, N(v) is the neighbor node, is the normalization coefficient, Weight matrix; learn the conservation and differences of metabolic pathways between species through GNN.

7. The small animal in vivo multimodal imaging system according to claim 1, characterized in that: The pharmacokinetic analysis module receives the registered images provided by the multimodal data fusion module, combines the multiscale PBPK model with tissue microstructure data, calculates the distribution and metabolic rate of the drug in heterogeneous tissues, obtains pharmacokinetic parameters, learns cross-species metabolic kinetics through the GNN-ODE algorithm, and uses the multi-species in vitro-in vivo joint verification platform to cross-validate the reliability of the model. The specific steps are as follows: The multimodal data fusion module outputs the registered image: The multimodal data fusion module outputs the registered image sequence, including the anatomical structure and metabolic distribution; Multiscale PBPK model construction and parameter initialization: The multiscale PBPK model divides the tissue into multiple compartments, each of which contains the following parameters: blood perfusion rate ; Tissue volume ; Drug partition coefficient : Parameter initialization formula: in, is the drug concentration in compartment i, is the metabolic rate; Tissue microstructure data integration: vascular density is measured by MRI perfusion imaging, cell type distribution is measured by tissue section staining and AI segmentation, and parameter correction is adjusted according to microstructure data and ; GNN-ODE algorithm learns cross-species metabolic dynamics: combining the PBPK model differential equation with graph neural network to learn cross-species metabolic laws. The differential equation form is: Where C(t) is the drug concentration vector, is the PBPK parameter, G is the cross-species metabolic profile; GNN embedding update: in, is the embedding of node v, N(v) is the neighboring node; Use the Runge-Kutta method to solve the differential equation and predict the change of drug concentration over time; Cross-validation of a multi-species in vitro-in vivo joint validation platform: drug metabolism rates are measured using a liver chip, PET / MRI data are collected in mouse / primate models, and pharmacokinetic parameters are calculated; Cross-validation: Input: in vitro chip data and in vivo animal data, calculate the loss function, including mean square error and Pearson correlation coefficient: ； Cross-species pharmacokinetic parameter output: Pharmacokinetic parameters: Volume of distribution , clearance rate CL, half-life ; Cross-species prediction: mapping of pharmacokinetic parameters from mouse to primate to human.

8. A small animal living multimodal imaging method corresponding to a small animal living multimodal imaging system according to claim 1, characterized in that: include: Animal model selection and preparation: Select a small animal model, anesthetize it, and secure it to the imaging platform; Tracer injection and dose optimization: The tracer dosage was calculated based on the dose prediction model and administered via tail vein injection; Synchronous multimodal image acquisition: Use photon counting detectors to collect fluorescence / bioluminescence signals, initiate PET / SPECT scans, combine deep learning dynamic sparse sampling algorithms to optimize data acquisition, execute MRI sequences, and obtain anatomical structure information; Real-time physiological parameter monitoring: Record physiological parameters through real-time dynamic monitoring module; Multimodal image registration and fusion: The original multimodal image is input into the Transformer cross-modal registration network to generate a registered image, which is then combined with the multi-species pharmacokinetic knowledge graph to construct a cross-modal metabolic joint map; Pharmacokinetic parameter quantification: The registered images are input into a multi-scale PBPK model to extract pharmacokinetic parameters. The model fit is optimized using the GNN-ODE algorithm, and the results are verified using a multi-species in vitro-in vivo joint validation platform. Real-time imaging optimization: Utilizes quantum computing OSEM algorithms to optimize PET reconstruction, shorten reconstruction time, and store analysis results in a multimodal data lake; Mining cross-species metabolic patterns: Analyzing metabolic differences among mouse, dog, primate, and humanized liver chips through a multi-species pharmacokinetic knowledge graph, and utilizing a federated learning framework to enable cross-center model collaborative training. Model validation and error analysis: Compare predicted pharmacokinetic parameters with measured values, calculate errors, and evaluate model reliability using a multi-species in vitro-in vivo combined validation platform.

Images