A method and system, device and apparatus for in vivo dosimetry of radiopharmaceuticals

Abstract

The application discloses a kind of radioactive medicine in-vivo metering method and system, device and equipment, it is related to nuclear medicine technical field.The method is carried out organ automatic segmentation using deep learning model, improve the degree of automation, and avoid the problem that there is subjective error in long time of existing hand-drawing, using Monte Carlo simulation method realizes patient individualized modeling, and tracks the energy deposition of each ray, considers the tissue heterogeneity, improves the cumulative dose calculation accuracy, selects the optimal fitting model most consistent with the corresponding organ pharmacokinetics based on AICc criterion, greatly improves the mathematical precision of final total absorbed dose.

Description

Technical Field

[0001] This application relates to the field of nuclear medicine technology, and in particular to a method, system, device and equipment for in vivo dosing of radiopharmaceuticals. Background Technology

[0002] In existing technologies, individualized dosimetric assessment for targeted radiopharmaceutical therapy (such as 177Lu-DOTATATE (177Lu-octreotide) for neuroendocrine tumors and 177Lu-PSMA (177Lu-tesivipitide injection) for prostate cancer) is crucial for determining treatment safety and efficacy. However, current routine clinical procedures and computational software have the following significant shortcomings: 1. Heavily reliant on manual labor, with extremely low clinical conversion rate: Accurate dose calculation requires acquiring multi-phase (usually 3-5 time points) whole-body large-field PET (Positron Emission Tomography) / SPECT (Single Photon Emission Computed Tomography) images. Currently, it mainly relies on physicians manually outlining organs at risk such as the kidneys, liver, and spleen, as well as tumor target areas, one phase at a time and one slice at a time. Manual outlining for a single patient usually takes more than 4 hours and involves significant inter-observer variability (i.e., subjective error), which is simply unacceptable for high-throughput routine clinical applications.

[0003] 2. Using a "standard phantom" instead of actual patient anatomy fails to achieve true individualization: Existing mainstream commercial dose calculation software (such as OLIDA / EXM (Organ Level Internal Dose Assessment / Exponential Modeling)) primarily relies on the "standard human" mathematical phantom from the MIRD (Medical Internal Radiation Dose Committee). This method assumes that the patient's organ size, relative position, and tissue density are fixed and standard, completely ignoring the patient's own body type, whether organs have been removed, or whether they are enlarged. For patients with extreme body types or abnormal anatomical structures, there is a significant gap between the dose calculated based on the standard phantom and the actual absorbed dose.

[0004] 3. The integral fitting model is too simplistic and ignores pharmacokinetic differences: The uptake and clearance patterns (biological half-life) of radiopharmaceuticals vary significantly across different organs. Most existing studies or software only provide single-exponential or double-exponential curve fitting, which cannot adaptively match the complex multiphase biological clearance process, resulting in an error of more than 20% in the integral of the area under the curve (AUC) of the time-activity curve (TAC). Summary of the Invention

[0005] The purpose of this application is to provide a method, system, apparatus, and equipment for in vivo radiopharmaceutical metrology, so as to improve the automation and accuracy of in vivo radiopharmaceutical metrology.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] In a first aspect, this application provides a method for in vivo dosing of a radiopharmaceutical, comprising: Acquire a multi-temporal medical image set of the patient; the medical image set includes emission computed tomography (ECT) images and CT (Computed Tomography) images of the same temporal phase; the emission computed tomography images are PET images and / or SPECT images; The target CT image is segmented into multiple organs using an organ segmentation model to obtain a segmentation mask for each organ; the organ segmentation model is obtained by training a deep learning model; the target CT image is any CT image from multiple temporal CT images. Monte Carlo simulations were performed on the medical image sets for each time phase to obtain three-dimensional dose rate distribution images for each time phase; The segmentation mask of each organ in the same time phase is registered with the three-dimensional dose rate distribution image to obtain the dose rate of each organ in each time phase. Based on the dose rate of each organ at each time phase, an optimal fitting model for each organ is constructed according to the AICc (Akaike Information Criterion corrected for small sample sizes) criterion; the optimal fitting model is a fitting model of time versus dose rate. The optimal fit model for each organ is integrated over time to obtain the cumulative absorbed dose for each organ.

[0008] Optionally, Monte Carlo simulations are performed on the medical image set for each time phase to obtain a three-dimensional dose rate distribution image for each time phase, which also includes: The third-order BSpline (Basis Spline) interpolation algorithm was used to resample the CT images of each time phase to make the spatial properties of the CT images and the emission tomography images of the same time phase consistent.

[0009] Optionally, the target CT image is segmented into multiple organs using an organ segmentation model to obtain a segmentation mask for each organ, followed by: The nearest neighbor interpolation algorithm is used to resample the segmentation mask of each organ based on the emission tomography images of multiple time phases, so that the spatial attributes of the segmentation mask and the emission tomography image of the same time phase are consistent.

[0010] Optionally, Monte Carlo simulation is performed on the medical image set for each time phase to obtain a three-dimensional dose rate distribution image for each time phase, specifically including: A density phantom of the patient is constructed based on CT images of phase k; k = 1, 2, ..., K; K is the total number of phases; The activity value of each pixel in the emission tomography image of phase k is assigned to the corresponding voxel of the density phantom to obtain the three-dimensional particle source distribution matrix. Monte Carlo simulation was performed based on the three-dimensional particle source distribution matrix to obtain a three-dimensional dose rate distribution image of phase k.

[0011] Optionally, the deep learning model is an improved U-Net model or a TotalSegmentator model. The improved U-Net model includes a three-dimensional U-Net (i.e., U-shaped convolutional neural network) base model, and spatial attention and channel attention mechanisms set at the skip connections of the three-dimensional U-Net base model.

[0012] Optionally, the loss function used during the training of the deep learning model is: ; in, For loss function, , , These are Dice loss, cross-entropy loss, and focus loss, respectively. , , These are the weight coefficients for Dice loss, cross-entropy loss, and focus loss, respectively.

[0013] Optionally, based on the dose rate of each organ at each time phase, an optimal fitting model for each organ is constructed according to the AICc criterion, specifically including: The dose rates of organ i at multiple time phases are respectively substituted into multiple time-dose rate models for curve fitting to obtain multiple fitting models for organ i; the multiple time-dose rate models include: single exponential model, double exponential model and trapezoidal integral model; i=1,2,...,I, where I is the number of organs segmented; Based on the AICc criterion, the optimal fitting model for organ i is selected from multiple fitting models for organ i.

[0014] Secondly, this application provides a radiopharmaceutical in vivo dosing system, including: a medical image input interface, a central processing unit, and a display; The central processing unit is connected to the medical image acquisition device through the medical image input interface, and the central processing unit is also connected to the display through the video interface; The central processing unit is used to obtain the cumulative absorbed dose of each organ using the above-described in vivo radiopharmaceutical dosing method.

[0015] Thirdly, this application provides an in vivo radiopharmaceutical dosing device, which applies the above-mentioned in vivo radiopharmaceutical dosing method, and the in vivo radiopharmaceutical dosing device includes: The image preprocessing module is used to acquire a multi-temporal medical image set of the patient; the medical image set includes emission computed tomography (ECT) images and CT images of the same temporal phase; the ECT images are PET images and / or SPECT images; The automatic organ segmentation module is used to segment multiple organs in a target CT image using an organ segmentation model to obtain a segmentation mask for each organ; the organ segmentation model is obtained by training a deep learning model; the target CT image is any CT image from multiple temporal CT images; The GATE dose simulation module is used to perform Monte Carlo simulations on medical image sets at each time point to obtain a three-dimensional dose rate distribution image at each time point. The time-dose rate curve statistics and integration module is used to register the segmentation mask of each organ and the three-dimensional dose rate distribution image at the same time phase to obtain the dose rate of each organ at each time phase. Based on the dose rate of each organ at each time phase, the optimal fitting model of each organ is constructed based on the AICc criterion. The optimal fitting model is a fitting model of time and dose rate. The optimal fitting model of each organ is integrated on the time axis to obtain the cumulative absorbed dose of each organ.

[0016] Fourthly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for in vivo dosing of radiopharmaceuticals.

[0017] According to the specific embodiments provided in this application, this application has the following technical effects.

[0018] This application provides a method, system, device, and equipment for in vivo radiopharmaceutical dosing. This application uses a deep learning model for automatic organ segmentation, which improves the degree of automation and avoids the problems of long time and subjective error in manual sketching. It uses the model Carlo simulation method to realize individualized patient modeling and tracks the energy deposition of each ray. It considers tissue inhomogeneity and improves the accuracy of cumulative dose calculation. Based on the AICc criterion, it selects the optimal fitting model that best fits the pharmacokinetics of the corresponding organ, which greatly improves the mathematical accuracy of the final total absorbed dose. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart of an in vivo radiopharmaceutical dosage method provided in an embodiment of this application.

[0021] Figure 2 This is a schematic diagram illustrating an in vivo dosing method for radiopharmaceuticals provided in an embodiment of this application.

[0022] Figure 3 This is a comparison image of a CT image before and after resampling, provided in an embodiment of this application.

[0023] Figure 4 This is a schematic diagram of the automatic segmentation result provided by TotalSegmentator according to an embodiment of this application.

[0024] Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] In one exemplary embodiment, a method for in vivo dosing of a radiopharmaceutical is provided, such as... Figure 1 and Figure 2 As shown, it includes the following steps 101-106.

[0028] Step 101: Acquire a multi-phase medical image set of the patient; the medical image set includes emission computed tomography (ECT) images and CT images of the same phase; the ECT images are PET images and / or SPECT images.

[0029] Step 102: Perform multi-organ segmentation on the target CT image using an organ segmentation model to obtain a segmentation mask for each organ; the organ segmentation model is obtained by training a deep learning model; the target CT image is any CT image from multiple temporal CT images.

[0030] Step 103: Perform Monte Carlo simulation on the medical image group for each time phase to obtain a three-dimensional dose rate distribution image for each time phase.

[0031] Step 104: Perform registration operations on the segmentation mask of each organ in the same time phase and the three-dimensional dose rate distribution image to obtain the dose rate of each organ in each time phase.

[0032] Step 105: Based on the dose rate of each organ in each time phase, construct the optimal fitting model for each organ according to the AICc criterion; the optimal fitting model is a fitting model of time and dose rate.

[0033] Step 106: Integrate the optimal fitting model for each organ on the time axis to obtain the cumulative absorbed dose for each organ.

[0034] Implementing steps 101-106 above can achieve the following technical effects.

[0035] This solution overcomes the bottleneck of time-consuming manual sketching by seamlessly integrating deep learning models with the GATE (Geant4 Application for Tomographic Emission, a medical imaging Monte Carlo simulation platform) Monte Carlo physics engine. Only the original image needs to be input to automatically complete the 3D segmentation of dozens of organs. The overall processing and computation time for a single patient has been reduced from "several days / hours" to <2 hours, achieving unattended "one-click" end-to-end computation. Specifically, it seamlessly integrates deep learning models (exemplary models such as the improved U-Net model and the TotalSegmentator model) with GATE Monte Carlo to achieve one-click individualized dose calculation.

[0036] Abandoning standard phantoms, true voxel-level individualized simulation is achieved: Completely discarding traditional MIRD standard phantoms, it directly generates realistic, personalized density matrices based on the patient's own CT images, combined with "piecewise linear HU-density mapping." By calling the GATE engine to track the energy deposition of each ray, and considering tissue inhomogeneity, the cumulative dose calculation error for critical limit organs such as the kidneys is reduced to <8%, far exceeding the accuracy of traditional commercial software.

[0037] To address the cumulative dose bias caused by rigid integral models, a "multi-model adaptive integration strategy" based on the information theory AICc criterion is proposed. Instead of blindly using a single function, the system lets the data speak for itself, automatically selecting the optimal simulation model from single-exponential, double-exponential, and trapezoidal integrals that best fits the pharmacokinetics of the organ, significantly improving the mathematical accuracy of the final total absorbed dose. In other words, the AICc-based multi-model adaptive integration strategy improves the accuracy of cumulative dose.

[0038] The aforementioned emission computed tomography images are PET images and / or SPECT images. In the following description, following the conventional practice in the medical field, PET / SPECT images are used to represent the emission computed tomography images.

[0039] In another exemplary embodiment, because the patient's position, respiratory status, and bowel fullness change at different time points during scanning, the CT (Computed Tomography) images and PET / SPECT images are spatially inconsistent. Existing conventional rigid registration or low-order interpolation methods are prone to causing blurred organ boundaries and volume deformation, resulting in the Standardized Uptake Value (SUV) being incorrectly assigned to adjacent tissues, causing a cumulative dose deviation as high as 10%-30%. In other words, conventional rigid registration or low-order interpolation methods have large multi-phase registration errors, leading to serious dose deviations.

[0040] To address this issue, this example includes the following step 201, preceding step 103 above.

[0041] Step 201: The third-order BSpline interpolation algorithm is used to resample the CT images of each time phase to make the spatial properties of the CT images and the emission tomography images of the same time phase consistent.

[0042] Step 201 specifically includes: extracting the geometric spatial information (including origin coordinates, voxel spacing, image size, and orientation matrix) of the corresponding temporal emission tomographic image, such as... Figure 3 As shown, Figure 3In the image, (a), (b), and (c) represent the reference image, the floating image, and the density-corrected CT image, respectively. During the interpolation process, the emission tomography image is used as the reference image (Fixed Image), as shown below. Figure 3 As shown in (a), the CT image is a moving image, as... Figure 3 As shown in (b), based on the SimpleITK (Simple Insight Segmentation and Registration Toolkit) library, a third-order BSpline (Basis Spline) interpolation algorithm is used to resample the CT image voxel by voxel, making its spatial properties completely consistent with the emission tomographic image, thus obtaining a density-corrected CT image, as shown in (b). Figure 3 As shown in (c) in the figure.

[0043] The purpose of step 201 is that third-order BSpline interpolation can smoothly and accurately preserve the continuous density variation characteristics of CT images. Resampling it to a space consistent with the functional images is to eliminate spatial differences between multimodal images.

[0044] In step 201, BSpline interpolation is preferentially used during resampling. This resampling step (i.e., step 201) is skipped when the CT image and PET / SPECT image spaces are completely identical. In this embodiment, a visual comparison image before and after resampling can also be generated. This step uses higher-order BSpline interpolation to significantly reduce registration errors.

[0045] This example also includes step 202 following step 102 above.

[0046] Step 202: Using the nearest neighbor interpolation algorithm, the segmentation mask of each organ is resampled based on the multi-temporal emission tomography images to make the spatial attributes of the segmentation mask and the emission tomography image of the same temporal phase consistent.

[0047] The specific implementation process of step 202 is as follows: traverse the multi-temporal sequence, obtain specific spatial metadata of the emission tomography image of each temporal phase, and use the Nearest-Neighbor Interpolation algorithm to project and resample the segmentation mask into the space of the emission tomography image of each temporal phase.

[0048] The purpose of step 202 is to segment the mask into discrete integer label data (e.g., 1 represents the liver and 2 represents the kidney). Nearest neighbor interpolation can strictly maintain the integer value of the label without changing it, avoiding meaningless artifact smoothing values ​​(e.g., 1.5), thereby ensuring the absolute accuracy of the mapping of organ anatomical boundaries at each time phase.

[0049] Based on steps 101-106 above, implementing steps 201-202 above can achieve the following technical effects.

[0050] The problem of multi-temporal deformation and registration error was solved by adopting the strategy of "single high-precision segmentation + third-order BSpline / nearest neighbor high-order interpolation resampling reuse". This avoids the boundary inconsistency caused by multiple segmentations, perfectly corrects the spatial misalignment of multi-temporal cross-modal images, and preserves the morphology and radioactivity fidelity of organs to the greatest extent.

[0051] In another exemplary embodiment, step 102 above is used to select any CT image and use a model segmentation model to perform automatic multi-organ segmentation to obtain a segmentation mask containing organs such as the liver, kidney, and spleen. The segmentation result is as follows: Figure 4 As shown, Figure 4 In the above, (a), (b), and (c) represent the CT image, the segmentation mask image, and the fused image, respectively. The fused image is obtained by superimposing the CT image and the segmentation mask image of the same region. The deep learning model in step 102 is a pre-trained improved U-Net model or a TotalSegmentator model. The improved U-Net model includes a 3D U-Net base model and spatial and channel attention mechanisms set at the skip connections of the 3D U-Net base model. This improved U-Net model introduces spatial and channel attention modules at the skip connections of the 3D U-Net base model, which can enhance the feature extraction ability of organs with complex shapes and small volumes (such as the spleen and renal cortex).

[0052] In this embodiment, data augmentation strategies such as elastic deformation, random rotation scaling, and adaptive histogram equalization are introduced during the training of the deep learning model to overcome pose variations caused by different patient body shapes and scanning bed positions.

[0053] In this embodiment of the application, the loss function used in the training process of the deep learning model is: ; in, For loss function, , , These are Dice loss, cross-entropy loss, and focus loss, respectively. , , These are the weight coefficients for Dice loss, cross-entropy loss, and focus loss, respectively. This loss function addresses the extreme class imbalance problem caused by the large size differences among multiple organs in the body by introducing Focal Loss, thereby improving the edge segmentation accuracy of small organs.

[0054] In another exemplary embodiment, the deep learning model used in step 102 is the TotalSegmentator series or a fine-tuned version thereof, and the training dataset consists of no less than 1,000 whole-body CT images, segmenting no less than 50 types of organs.

[0055] In another exemplary embodiment, step 103 above first performs decay correction and SUV normalization on each phase PET / SPECT image, and calculates the average (SUVmean), maximum (SUVmax), and peak (SUVpeak) absolute values ​​of radioactivity in the segmentation mask of each organ to generate a time-activity curve. Here, SUVmean represents the average value of all pixels within the region of interest, which better represents the overall metabolic level, but is affected by the delineation range; SUVmax represents the value of the highest pixel within the region of interest, is the most commonly used, is less affected by the measurement method, and has good repeatability; SUVpeak represents the metabolically most active area within the region of interest (typically 1 cm). 3 The average value of the former two is combined with the advantages of the latter. In this embodiment, the region of interest is the corresponding region of the segmented organ.

[0056] The specific process of step 103 is as follows: the resampled multi-temporal segmentation mask is superimposed on its corresponding standardized PET / SPECT image for mask operation, and the absolute value of radioactivity in the region of interest (ROI) of each organ is extracted and counted (the mean value is obtained by calculating the mean activity value SUVmean, the maximum value is obtained by calculating the maximum activity value SUVmax, the peak activity value SUVpeak, etc. are determined). Then, the activity values ​​(mean activity value, maximum activity value or peak activity) extracted from each organ are sequence matched with their corresponding scanning acquisition time points (such as 0.5h, 4h, 24h, etc.) to fit and draw the time-activity data point set of each organ.

[0057] The time-activity curve obtained in step 103 above reflects the dynamic uptake and metabolic clearance of radiopharmaceuticals in different organs (pharmacokinetics), providing the core mathematical basis for the integral calculation of total decay times (Time-Integrated Activity, TIA) in step 106 above, and is the key preliminary data for determining the final absorbed dose.

[0058] In another exemplary embodiment, step 104 above is based on GATE Monte Carlo software, which automatically generates a material database, density table, source distribution (ACTM.mhd), and macro file for each time phase to achieve time-phase dose rate simulation. Step 103 above can be replaced by steps 301-303 below.

[0059] Step 301: Construct a density phantom of the patient based on the CT images of phase k; k=1,2,...,K; K is the total number of phases.

[0060] Using automated scripts, CT images or density-corrected CT images (CT images resampled in step 201) are converted into a voxelized material density matrix, i.e., a density phantom, that can be recognized by GATE, according to a piecewise linear mapping relationship. This piecewise linear mapping relationship characterizes the conversion relationship between the HU (Hunsfield Unit) value and density of the CT image, where density refers to physical density (unit: g / cm³). 3 (or electron density (ratio relative to water)).

[0061] Step 302: Assign the activity value of each pixel in the emission tomography image of phase k to the corresponding voxel of the density phantom to obtain the three-dimensional particle source distribution matrix. Step 302 is used to convert the PET / SPECT image of phase k into a three-dimensional particle source distribution matrix.

[0062] Step 303: Perform Monte Carlo simulation based on the density phantom and the three-dimensional particle source distribution matrix to obtain a three-dimensional dose rate distribution image of phase k.

[0063] Automatically generate GATE macro files to define the decay energy spectrum, physical decay model, cross-section database, and number of tracking particles (Histories) for radioactive source nuclides (such as 177Lu). Submit the macro file to the GATE engine to run the Monte Carlo radiation transfer simulation, track the energy deposition of rays in the density phantom, and finally output a three-dimensional dose rate distribution image of phase k.

[0064] Traditional single-scale conversion leads to severe distortion in voxel density assignment for extreme densities (such as extremely low-density inflated lung lobes and extremely high-density dense bone), resulting in errors in the calculation of radiation scattering and absorption at bone-soft tissue interfaces or lung interfaces. In step 103 above, piecewise linear mapping is used to generate a high-precision three-dimensional physical density map based 100% on the patient's actual personalized anatomical structure (rather than a standard phantom). This significantly reduces Monte Carlo dose simulation errors of radioisotopes in the lungs, bones, and adjacent organs, achieving truly "voxel-level" precise individualized internal radiation dose calculation.

[0065] In step 104 above, the GATE simulation, based on the patient's own CT images or density-corrected CT images, employs a piecewise linear mapping relationship to significantly reduce lung and bone density errors. Specific explanation of the piecewise linear mapping relationship follows: In Monte Carlo dose simulations (such as GATE), accurate material density and composition are core prerequisites for calculating radiation energy deposition (i.e., absorbed dose). Due to the vast differences in tissue density within the human body, the CT value (i.e., HU value) on CT images differs significantly from the actual physical density (g / cm³). 3 The relationship between them is not a simple linear one.

[0066] This application employs a piecewise linear mapping relationship to divide the HU value range of a patient's CT image into multiple typical tissue ranges (e.g., air-lung tissue segment, fat-soft tissue / water segment, spongy bone-cortical bone segment, etc.). Within each specific HU range, an independent linear equation is used ( ,in, For physical density, (where HU is the value of the voxel and a and b are coefficients of the linear equation) precisely maps the HU value of the voxel to the corresponding physical density and assigns corresponding material properties (such as different effective atomic numbers).

[0067] In another exemplary embodiment, step 104 above registers the three-dimensional dose rate distribution image obtained in step 103 with the segmentation masks of each organ, and statistically obtains the time-dose rate curves of each organ. Specifically, the three-dimensional dose rate distribution image obtained in step 103 is read into the system, and spatial extraction is performed on the three-dimensional dose rate distribution image using the precisely aligned segmentation masks of each organ. The average dose rate (Gy / h or mGy / MBq) of voxels within the mask region of each organ is calculated, and the average dose rate values ​​of multiple time phases are sorted according to their actual scanning time points (t1, t2, ..., t...). K The data are arranged and combined to construct independent "time-dose rate" scatter plots for each organ, and then time-dose rate curves are generated.

[0068] In another exemplary embodiment, in steps 105 and 106 above, the time-dose rate curves are fitted with single-exponential, double-exponential, and trapezoidal integrals, respectively. The optimal model is automatically selected using the AICc (Akaike Information Criterion corrected for small sample sizes) criterion, and the cumulative absorbed dose of each organ is calculated.

[0069] For each fitted model, its AICc value is calculated using the following formula: ; in, The number of parameters for the fitted model; specifically, for a single exponential model: For the double exponential model: For the trapezoidal integral model: ; L is the maximum likelihood function value of the fitted model; n is the number of data points, i.e., the number of phase points.

[0070] Compare the AICc values ​​of all fitted models and select the model with the smallest AICc value as the optimal fitted model. This model can fit the data well without overfitting due to too many parameters. Moreover, the smaller the AICc value, the better the model achieves a balance between fitting accuracy and model complexity.

[0071] The specific process is as follows: Using the nonlinear least squares method, the extracted time-dose rate scatter points are substituted into both a single exponential decay model and a double exponential (uptake + clearance) model for curve fitting, while simultaneously calculating the trapezoidal integral (with a purely physical decay tail at the end). For the exponential fitting model, its modified Akaike Information Criterion (AICc) score is calculated. The system automatically compares the scores and selects the fitting model with the smallest AICc value (i.e., the model complexity and fit are optimally balanced) as the optimal model. Finally, the area under the bounds (AUC) of this optimal fitting model is calculated on the time axis [0, ∞) to obtain the final cumulative absorbed dose of the organ.

[0072] In another exemplary embodiment, step 105 described above can be replaced by steps 401-402.

[0073] Step 401: Substitute the dose rates of organ i at multiple time phases into multiple time-dose rate models for curve fitting to obtain multiple fitting models for organ i; the multiple time-dose rate models include: single exponential model, double exponential model and trapezoidal integral model; i=1,2,...,I, where I is the number of organs segmented.

[0074] Step 402: Select the optimal fitting model for organ i from multiple fitting models based on the AICc criterion.

[0075] In another exemplary embodiment, the formula for calculating the cumulative absorbed dose in step 106 above is: ; in, This represents the area under the curve of the optimally fitted model. For cumulative absorbed dose, Let be the dose rate at time t. The time range of this integral is from the injection time to a sufficiently long time to ensure that the complete radioactive decay process is captured.

[0076] The optimal fitting model was calculated using the following three methods and automatically selected by AICc. When the optimal fitting model is different... The calculation methods are different, as detailed below.

[0077] a) Single index: ; ; in, λ: decay constant; Cumulative absorbed dose, in units of Bq·s or Bq·h; Initial activity, in Bq; Physical half-life, measured in seconds (s), hours (h), or days (d). : Natural logarithm constant (approximately 0.693).

[0078] b) Double exponent: ; ; in, Fast phase clearing rate; Slow phase clearance rate; Fast phase elimination half-life; : Slow phase elimination half-life.

[0079] c) Trapezoidal integral + physical decay tail correction.

[0080] ; ; ; in, This is the cumulative amount obtained by the trapezoidal integral. This is the cumulative amount obtained by correcting for the physical decay tail. This represents the total number of time phases. , These are the measurement times corresponding to phase k and phase k+1, respectively. , , They are respectively time, time, The dose rate at any given time.

[0081] In another exemplary embodiment, after step 106 above, the method further includes: summing up the dose rates of all phases and overlaying them with the segmentation mask to generate a dose-volume histogram (DVH) and a dose color overlay map.

[0082] The method implementation of this application is based on Python (Pyson, a computer programming language) + SimpleITK (Simple Insight Segmentation and Registration Toolkit), PyTorch (PyTorch, an open-source deep learning framework), and the GATE framework to achieve full automation. It employs a core strategy of "single-time deep learning segmentation + multi-temporal multiplexing + patient-specific density-corrected CT," combined with GATE Monte Carlo simulation and multi-model time integration, to achieve end-to-end automated calculation from image to cumulative dose. The single-case calculation time is <2 hours, and the kidney dose error is <8%. The single-case multi-temporal (6 phases) calculation time does not exceed 2 hours. Python serves as the main control program, coordinating the other three tools (SimpleITK + PyTorch + GATE) and handling data flow, file I / O, and logic control. SimpleITK handles medical image reading and writing, geometric transformations, and resampling. PyTorch runs deep learning models to achieve automatic organ and tumor segmentation. GATE performs physics-based particle transport simulations to calculate energy deposition.

[0083] In another exemplary embodiment, another implementation of the above-described cumulative absorbed dose is provided, as follows: The segmentation mask of each organ in each time phase is superimposed with the emission tomography image of the corresponding time phase to perform mask operation, obtain the activity value of each organ in each time phase, and generate the time-activity curve of each organ. Based on the time-activity curves of each organ at each time phase, calculate the total number of decays (TIAs) for each organ at each time phase: ,in, For each organ at any given time phase, the time-activity curves are given, where t is the time variable. Based on the time-activity curves of each organ at each time phase, the average dose rate coefficient S is calculated: ;in, Let be the dose rate at time t; Calculate the final cumulative dose based on the total decay number (TIA) and the average dose rate coefficient (S): .

[0084] Similarly, in this embodiment, the integration time range is from the injection time to a sufficiently long time to ensure that the complete radioactive decay process is captured.

[0085] The above-described method implementations can be implemented based on computer programs, are fully open source and reproducible, and have extremely low costs.

[0086] In one exemplary embodiment, a radiopharmaceutical in vivo dosing system is provided, comprising: a medical image input interface, a central processing unit, and a display; The central processing unit is connected to the medical image acquisition device through the medical image input interface, and the central processing unit is also connected to the display through the video interface; The central processing unit is used to obtain the cumulative absorbed dose of each organ using the above-described in vivo radiopharmaceutical dosing method.

[0087] In another exemplary embodiment, the aforementioned medical image input interface can be a DP interface, an HDMI interface, or a USB-C interface. The aforementioned video interface can be an HDMI interface or a DP interface. The aforementioned central processing unit has at least image processing and computing functions, and a computer is preferred.

[0088] Based on the same inventive concept, this application also provides an in vivo radiopharmaceutical dosing device for implementing the aforementioned in vivo radiopharmaceutical dosing method. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations of one or more in vivo radiopharmaceutical dosing device embodiments provided below can be found in the limitations of the in vivo radiopharmaceutical dosing method described above, and will not be repeated here.

[0089] In one exemplary embodiment, an in vivo radiopharmaceutical dosing device is provided, comprising: The image preprocessing module is used to acquire a multi-temporal medical image set of the patient; the medical image set includes emission computed tomography (ECT) images and CT images of the same temporal phase; the ECT images are PET images and / or SPECT images; The automatic organ segmentation module is used to segment multiple organs in a target CT image using an organ segmentation model to obtain a segmentation mask for each organ; the organ segmentation model is obtained by training a deep learning model; the target CT image is any CT image from multiple temporal CT images; The GATE dose simulation module is used to perform Monte Carlo simulations on medical image sets at each time point to obtain a three-dimensional dose rate distribution image at each time point. The time-dose rate curve statistics and integration module is used to register the segmentation mask of each organ and the three-dimensional dose rate distribution image at the same time phase to obtain the dose rate of each organ at each time phase. Based on the dose rate of each organ at each time phase, the optimal fitting model of each organ is constructed based on the AICc criterion. The optimal fitting model is a fitting model of time and dose rate. The optimal fitting model of each organ is integrated on the time axis to obtain the cumulative absorbed dose of each organ.

[0090] In another exemplary embodiment, the above-described apparatus further includes a visualization module for generating a comparison image before and after resampling, a segmentation mask overlay image, a dose color overlay image, and a time-activity curve fitting image.

[0091] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for in vivo dosing of radiopharmaceuticals.

[0092] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0093] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0094] The following specific examples are provided to illustrate the effectiveness of this solution.

[0095] Example 1: 177Lu Treatment Case: The patient was a 68-year-old male who received 177Lu-DOTATATE treatment. Six time-phase PET / CT scans were collected at 0.5h, 4h, 24h, 48h, 96h, and 168h.

[0096] All CT images were resampled to PET space using BSpline interpolation, with a pixel size error of <0.1mm.

[0097] The 24-hour CT images were segmented using the TotalSegmentator model (Dataset297), with a Dice coefficient > 0.92.

[0098] The cumulative absorbed doses for each organ were obtained: 19.3±1.6 Gy for the left kidney, 18.7±1.8 Gy for the right kidney, and 28.5±2.1 Gy for the liver. Compared with OlinDA software, the dose deviation for the kidneys (including both the left and right kidneys) was only 6.8%, significantly better than the 20-30% error reported in the literature.

[0099] Example 2: For SPECT 99mTc cases, the same procedure applies, except that the relevant codes for PET images are replaced with those for SPECT images.

[0100] The above examples demonstrate that this application can operate stably under various radionuclides and treatment regimens in PET and SPECT, exhibiting strong versatility and promising clinical translation prospects.

[0101] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0102] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0103] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for in vivo dosing of a radiopharmaceutical, characterized in that, include: Acquire a multi-temporal medical image set of the patient; the medical image set includes emission computed tomography (ECT) images and CT images of the same temporal phase; the ECT images are PET images and / or SPECT images; The target CT image is segmented into multiple organs using an organ segmentation model to obtain a segmentation mask for each organ; the organ segmentation model is obtained by training a deep learning model; the target CT image is any CT image from multiple temporal CT images. Monte Carlo simulations were performed on the medical image sets for each time phase to obtain three-dimensional dose rate distribution images for each time phase; The segmentation mask of each organ in the same time phase is registered with the three-dimensional dose rate distribution image to obtain the dose rate of each organ in each time phase. Based on the dose rate of each organ at each time phase, an optimal fitting model for each organ is constructed according to the AICc criterion; the optimal fitting model is a fitting model of time versus dose rate. The optimal fit model for each organ is integrated over time to obtain the cumulative absorbed dose for each organ. Based on the dose rate of each organ at each time phase, an optimal fitting model for each organ is constructed according to the AICc criterion, specifically including: The dose rates of organ i at multiple time phases are respectively substituted into multiple time-dose rate models for curve fitting to obtain multiple fitting models for organ i; the multiple time-dose rate models include: single exponential model, double exponential model and trapezoidal integral model; i=1,2,...,I, where I is the number of organs segmented; The optimal fitting model for organ i is selected from multiple fitting models based on the AICc criterion. Monte Carlo simulations were performed on the medical image sets for each time phase to obtain three-dimensional dose rate distribution images for each time phase, specifically including: A density phantom of the patient is constructed based on CT images of phase k; k = 1, 2, ..., K; K is the total number of phases; The activity value of each pixel in the emission tomography image of phase k is assigned to the corresponding voxel of the density phantom to obtain the three-dimensional particle source distribution matrix. Monte Carlo simulations were performed based on the density phantom and the three-dimensional particle source distribution matrix to obtain a three-dimensional dose rate distribution image for phase k.

2. The in vivo dosage method for radiopharmaceuticals according to claim 1, characterized in that, Monte Carlo simulations were performed on the medical image sets for each time phase to obtain three-dimensional dose rate distribution images for each time phase. This process also included: The third-order BSpline interpolation algorithm was used to resample the CT images of each time phase to make the spatial properties of the CT images and the emission tomography images of the same time phase consistent.

3. The in vivo dosage method for radiopharmaceuticals according to claim 1, characterized in that, The target CT image is segmented into multiple organs using an organ segmentation model to obtain a segmentation mask for each organ. This process also includes: The nearest neighbor interpolation algorithm is used to resample the segmentation mask of each organ based on the emission tomography images of multiple time phases, so that the spatial attributes of the segmentation mask and the emission tomography image of the same time phase are consistent.

4. The in vivo dosage method for radiopharmaceuticals according to claim 1, characterized in that, The deep learning model is an improved U-Net model or a TotalSegmentator model. The improved U-Net model includes a three-dimensional U-Net base model and spatial attention and channel attention mechanisms set at the skip connections of the three-dimensional U-Net base model.

5. The in vivo dosage method for radiopharmaceuticals according to claim 1, characterized in that, The loss function used in training a deep learning model is: ； in, For loss function, , , These are Dice loss, cross-entropy loss, and focus loss, respectively. , , These are the weight coefficients for Dice loss, cross-entropy loss, and focus loss, respectively.

6. An in vivo radiopharmaceutical dosing system, characterized in that, include: Medical image input interface, central processing unit and display; The central processing unit is connected to the medical image acquisition device through the medical image input interface, and the central processing unit is also connected to the display through the video interface; The central processing unit is used to obtain the cumulative absorbed dose of each organ using the in vivo radiopharmaceutical dosing method according to any one of claims 1-5.

7. An in vivo dosing device for radiopharmaceuticals, characterized in that, The in vivo radiopharmaceutical dosing device uses the in vivo radiopharmaceutical dosing method according to any one of claims 1-5, and the in vivo radiopharmaceutical dosing device comprises: The image preprocessing module is used to acquire a multi-temporal medical image set of the patient; the medical image set includes emission computed tomography (ECT) images and CT images of the same temporal phase; the ECT images are PET images and / or SPECT images; The automatic organ segmentation module is used to segment multiple organs in a target CT image using an organ segmentation model to obtain a segmentation mask for each organ; the organ segmentation model is obtained by training a deep learning model; the target CT image is any CT image from multiple temporal CT images; The GATE dose simulation module is used to perform Monte Carlo simulations on medical image sets at each time point to obtain a three-dimensional dose rate distribution image at each time point. The time-dose rate curve statistics and integration module is used to register the segmentation mask of each organ and the three-dimensional dose rate distribution image at the same time phase to obtain the dose rate of each organ at each time phase. Based on the dose rate of each organ at each time phase, the optimal fitting model of each organ is constructed based on the AICc criterion. The optimal fitting model is a fitting model of time and dose rate. The optimal fitting model of each organ is integrated on the time axis to obtain the cumulative absorbed dose of each organ.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the in vivo dosing method for radiopharmaceuticals according to any one of claims 1-5.

Images