Tumor radiotherapy dose calibration method

Abstract

The invention discloses a tumor radiotherapy dose calibration method, and relates to the technical field of tumor radiotherapy, and the method comprises the steps: collecting multi-modal data in real time through 4D-CT, MRI and the like, extracting dynamic features through space-time registration, calculating a risk organ dose ratio and a tumor target dose deviation rate, and constructing a comprehensive evaluation function. The method has the advantages that the multi-modal dynamic data of the tumor target area are synchronously collected in real time, the space-time joint registration algorithm and the time sequence weight factor are combined, the limitation that existing static calibration depends on fixed parameters is broken through, and the method is suitable for the multi-modal dynamic data of the tumor target area. According to the method, 95% dose coverage of a tumor target region is guaranteed, and meanwhile, organ-endangering peak value receiving amount is controlled within a certain range of a clinical tolerance threshold value, so that the risk of excessive irradiation of normal tissues is reduced, and the radiotherapy dose calibration precision is prompted based on the deviation rate of a traditional method.

Description

Technical Field

[0001] The present invention relates to the technical field of tumor radiotherapy, and in particular to a tumor radiotherapy dose calibration method. Background Art

[0002] Tumor radiotherapy is an important method of killing tumor cells using ionizing radiation. Its efficacy is closely related to the accuracy of the dose. With the development of precision radiotherapy technology, how to dynamically adapt to the anatomical changes of the tumor target area and organs at risk during treatment has become a key challenge to improve the safety and effectiveness of radiotherapy.

[0003] Existing technologies have certain defects. First, existing technologies rely on fixed parameters for radiotherapy dose calibration, which makes it difficult to respond to tumor shrinkage and organ movement in a timely manner, easily leading to dose deviation and excessive irradiation of normal tissue. Second, traditional methods lack a "data acquisition-evaluation-calibration-iteration" closed-loop system and cannot dynamically adapt to individual anatomical changes. The calibration strategy is fixed and the optimization capability is insufficient. Therefore, we propose a tumor radiotherapy dose calibration method. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for calibrating tumor radiotherapy dose.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for calibrating a tumor radiotherapy dose, the calibration method comprising the following specific steps:

[0006] Step 1: Use 4D-CT, MRI, EPID, and PET / CT to synchronously collect the three-dimensional motion trajectory of the patient's tumor target area and organs at risk, soft tissue deformation, two-dimensional dose distribution within the radiation field, and tumor FDG metabolic activity data in real time to form a four-dimensional dynamic data set;

[0007] Step 2: Use the spatiotemporal joint registration algorithm to perform elastic registration of 4D-CT and MRI data, and introduce the time series weight factor λ t and the spatial deformation penalty term Ω(μ), the registered three-dimensional structure data is obtained, and the volume change rate of the tumor target area, the real-time displacement vector of the organ at risk and the dose hotspot coordinates are extracted;

[0008] Step 3: Using the real-time displacement vector of the organ at risk to correct the initial coordinates of the organ at risk, mapping the two-dimensional dose distribution data within the radiation field to the corrected organ region, calculating the ratio of the peak dose of the organ at risk to the clinical preset tolerance threshold, calculating the dose deviation rate between the actual 95% dose coverage value of the tumor target area and the preset planned dose, introducing a neural network model to predict the tumor regression correction factor, constructing a comprehensive evaluation function, and calculating the comprehensive evaluation value based on the ratio, dose deviation rate, and tumor regression correction factor;

[0009] Step 4: Perform dual-mode calibration based on the comprehensive evaluation value. When the tumor target dose deviation rate is greater than the set threshold and the OAR dose ratio is less than the set threshold, the dose adjustment factor is calculated based on the tumor target volume change rate, and the MLC leaf fine-tuning parameters are generated based on the tumor target displacement vector. When any OAR dose ratio is greater than or equal to the set threshold, the field angle is solved using a constrained optimization algorithm.

[0010] Step 5: Transmit the generated calibration parameters to the linear accelerator, perform field pre-verification based on EPID real-time dose data, and determine whether to perform calibration based on the gamma pass rate;

[0011] Step 6: Store the data from the calibration process into the database. When the data volume reaches the preset amount, perform incremental training on the LSTM neural network to update the tumor regression-dose correction mapping model and the risk organ dose prediction model.

[0012] As a further solution of the present invention: in step 1, a 4D-CT scanner integrated with the radiotherapy equipment is used to continuously capture the three-dimensional motion trajectory of the tumor target area and adjacent organs at risk during the patient's respiratory cycle at a frequency of 50ms / frame-60ms / frame, and the MRI guidance system is synchronously triggered to acquire T1-weighted images every 2min-3min to capture soft tissue deformation. An electronic portal imaging device is used to collect two-dimensional dose distribution data within the portal in real time at a frequency of 10Hz-20Hz, with an accuracy controlled within ±0.1Gy / mm 2 Before treatment, FDG metabolic activity data of tumors are obtained through PET / CT. All data are transmitted to the calibration control module of the treatment planning system in real time to form a four-dimensional dynamic data set including anatomical structure, functional metabolism, motion trajectory and real-time dose, which is recorded as D = {D CT ,D MRI ,D EPID ,D PET}, where D CT Contains the three-dimensional coordinates of the tumor target area and organs in each frame, D EPID is the real-time dose matrix.

[0013] As a further solution of the present invention: in the step 2, in the step 2, for the collected 4D-CT-D CT With MRI-D MRI Data, the improved spatiotemporal joint registration algorithm is used for elastic registration, and the time series weight factor λ is introduced t And the spatial deformation penalty term Ω(μ) is used to suppress respiratory motion artifacts. The time series weight factor expression is as follows:

[0014] λ t =0.9 (Δt / 300)

[0015] Where Δt is the time interval between the current frame and the initial frame, and the registration energy function is defined as:

[0016]

[0017] Among them, E sim is the similarity measure of the grayscale mutual information between 4D-CT and MRI, μ t is the displacement field at time t, α = 0.6, β = 0.3, γ = 0.1 are weight coefficients, T represents the total number of frames in the time series, and Ω (μ) is the sum of squares of the second-order derivatives of the displacement field;

[0018] After registration, the tumor target volume change rate V is extracted from the four-dimensional dynamic data set D rate , Real-time displacement vector of organs at risk and EPID real-time dose hotspot coordinates (x i ,y i ,z i ,D i ), D i represents the actual measured dose value of the dose hotspot area, let D i >1.1D plan For the over-punctual point, D plan It represents the planned dose preset in the treatment plan. The formula for calculating the tumor target volume change rate is as follows:

[0019]

[0020] Among them, V initial is the volume of the first treatment tumor target collected in step 1, V current is the tumor target volume during current treatment.

[0021] As a further solution of the present invention: In step 3, the displacement vector of step 2 is used Perform real-time correction on the initial coordinates (x0, y0, z0) of the organ at risk to obtain dynamic anatomical coordinates The EPID dose matrix D in step 1 is converted into EPID Map to the corrected organ area, calculate voxel dose values ​​and extract peak dose The calculation formula is as follows:

[0022]

[0023] Definition of organ-at-risk dose ratio η org , η org The calculation formula is as follows:

[0024]

[0025] Where v represents the actual region of the organ at risk (OAR) real For each voxel in the image, D(v) represents the actual region of the organ at risk (OAR). real The actual dose value of the voxel v, Preset tolerance threshold for clinical practice and calculate the actual 95% dose coverage value for the tumor target area Combined with planned dose D plan Calculate the dose deviation rate δ D , δ D The calculation is defined as follows:

[0026]

[0027] The LSTM neural network model is introduced. The LSTM model contains 3 hidden layers, 128 neurons in each layer, and the input features include the tumor target volume change rate, displacement vector and historical calibration parameters. rate 、 and historical calibration parameters to predict tumor regression correction factor f(V rate ), construct a comprehensive evaluation function to calculate the evaluation result S, the comprehensive evaluation function is as follows:

[0028] S=0.6δ D +0.3η org +0.1|f(V rate )|

[0029] When S≤0.15, the subsequent calibration process is not triggered, otherwise the subsequent calibration process is triggered.

[0030] As a further solution of the present invention: in step 4, according to the evaluation value S and the deviation type of step 3, a dual-mode calibration strategy is executed:

[0031] When δ D >10% and all η org When the dose compensation mode of the tumor target is activated, the dose compensation mode of the tumor target is activated based on V rate Calculate the dose adjustment factor k using the following formula:

[0032] k=1+0.6×|V rate |

[0033] The value range of k is 1.05≤k≤1.20. When the k value exceeds 1.20, it is 1.20. When the k value is less than 1.05, it is 1.05. The displacement vector of the tumor target area is used as the displacement vector. Generate MLC blade position adjustment parameter Δx j and Δy j , Δx j and Δy j The calculation is as follows:

[0034]

[0035] in, Field plane unit vector;

[0036] When all η org When the value is ≥0.8, the organ protection mode is activated and the radiation field angle θ is solved by the constrained gradient descent algorithm. new , define the objective function F as follows:

[0037]

[0038] Among them, D tol (v) represents the actual region of the organ at risk (OAR) real The clinical tolerance dose threshold of voxel v is constrained by |θ new -θ old |≤15° and the field path avoids the high-density risk area identified by step 2, θ old Indicates the initial angle of the current irradiation field.

[0039] As a further solution of the present invention: in step 5, the calibration parameters generated in step 4 are transmitted to the linear accelerator, and the calibration parameters include the dose factor k, the blade coordinates (x j +Δx j ,y j +Δy j ) and field angle θ new , based on the EPID data of step one, the field verification is performed, the dose distribution of the verification field is collected, and the γ pass rate of the planned dose after calibration is calculated. When γ ≥ 95%, calibration is performed, and then according to the δ of step three D Start the backup mechanism, δ D When the value is less than 5%, the historical parameters in the calibration database are called. D When ≥5%, it triggers a reminder for the physiotherapist to review.

[0040] As a further solution of the present invention: in step 6, after the calibration is completed, the original data of step 1, the registration parameters of step 2, the evaluation results of step 3, the calibration parameters of step 4 and the verification results of step 5 are stored in the calibration database. When the data accumulates to 50-60 cases, V rate With f(V rate ) as training samples, perform incremental training on the LSTM model in step 3, and update the tumor regression-dose correction mapping relationship and the risk organ dose prediction model.

[0041] By adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. The present invention uses 4D-CT, MRI, EPID and PET / CT to synchronously collect multimodal dynamic data such as the three-dimensional motion trajectory of the tumor target area, soft tissue deformation, real-time dose and tumor metabolic activity, and combines the spatiotemporal joint registration algorithm with the time series weight factor to accurately extract the volume change rate of the tumor target area and the real-time displacement vector of the organ at risk, providing real-time anatomical and functional metabolic data support for dynamic dose calibration. The LSTM neural network is used to construct a comprehensive evaluation function, which integrates and quantifies the tumor target area dose deviation rate, the organ at risk dose ratio and the tumor shrinkage biological effect correction factor, and intelligently judges the calibration trigger conditions and deviation types. For the problems of insufficient tumor target area dose and excessive organ at risk dose, the dual-mode The calibration strategy dynamically adjusts the dose intensity, MLC leaf position and field angle - the dose adjustment factor is calculated based on the volume change rate of the tumor target area and the leaf position is corrected. The field angle that avoids high-risk areas is solved through a constrained optimization algorithm to achieve accurate response to dynamic anatomical changes. This method breaks through the limitations of existing static calibration that relies on fixed parameters, and corrects dose deviations caused by tumor shrinkage and organ movement in real time. While ensuring 95% dose coverage of the tumor target area, the peak dose of the endangered organs is controlled within a certain range of the clinical tolerance threshold, reducing the risk of over-irradiation of normal tissues, and enabling the deviation rate of radiotherapy dose calibration accuracy based on traditional methods to be prompted, effectively improving the individualized accuracy and safety of tumor radiotherapy.

[0043] 2. The present invention realizes dynamic and precise control of tumor radiotherapy dose by establishing an adaptive closed-loop system of "data acquisition-deviation assessment-intelligent calibration-model iteration". First, 4D-CT, MRI and other equipment collect the motion trajectory of the tumor target area, soft tissue deformation, real-time dose and metabolic activity data in real time. After processing by the spatiotemporal registration algorithm, the volume change of the tumor target area and the displacement information of the organ are accurately obtained, providing real-time dynamic data support for closed-loop control. Secondly, the LSTM neural network integrates multi-dimensional information to generate a comprehensive evaluation value, intelligently identify the dose deviation of the tumor target area and the organ dose risk, and trigger the dual-mode calibration strategy - dynamically adjust the dose intensity and MLC blade position according to the volume change of the tumor target area, targeting The field angle is solved for organ displacement through a constrained optimization algorithm to achieve real-time dynamic correction of the dose distribution during treatment. The calibration process data is automatically imported into the database. When it accumulates to a preset scale, the training model is automatically incremented to update the tumor shrinkage-dose correction mapping relationship, forming a self-improving closed loop of "data-driven-intelligent decision-making-dynamic adjustment-model optimization". This method realizes the adaptive response of radiotherapy dose calibration to the dynamic changes of individual anatomical structures through real-time interaction of multimodal data, precise decision-making of intelligent algorithms and iterative optimization of models, effectively reducing the dose to normal tissue while ensuring the accuracy of tumor target area dose coverage, and providing a closed-loop control solution for precision radiotherapy that integrates real-time monitoring, intelligent calibration and continuous optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the method steps in an embodiment of the present invention;

[0045] Figure 2 Schematic diagram of parameter comparison in embodiments of the present invention. DETAILED DESCRIPTION

[0046] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0047] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0048] Please see the attached Figure 1 -Attached Figure 2 The present invention provides a method for calibrating a tumor radiotherapy dose, the calibration method comprising the following specific steps:

[0049] Step 1: Use 4D-CT, MRI, EPID, and PET / CT to synchronously collect the three-dimensional motion trajectory of the patient's tumor target area and organs at risk, soft tissue deformation, two-dimensional dose distribution within the radiation field, and tumor FDG metabolic activity data in real time to form a four-dimensional dynamic data set;

[0050] Step 2: Use the spatiotemporal joint registration algorithm to perform elastic registration of 4D-CT and MRI data, and introduce the time series weight factor λ t and the spatial deformation penalty term Ω(μ), the registered three-dimensional structure data is obtained, and the volume change rate of the tumor target area, the real-time displacement vector of the organ at risk and the dose hotspot coordinates are extracted;

[0051] Step 3: Use the real-time displacement vector of the organ at risk to correct the initial coordinates of the organ at risk, map the two-dimensional dose distribution data within the radiation field to the corrected organ region, calculate the ratio of the peak dose of the organ at risk to the clinically preset tolerance threshold, calculate the dose deviation rate between the actual 95% dose coverage value of the tumor target and the preset planned dose, introduce a neural network model to predict the tumor regression correction factor, construct a comprehensive evaluation function, and calculate the comprehensive evaluation value based on the ratio, dose deviation rate, and tumor regression correction factor;

[0052] Step 4: Perform dual-mode calibration based on the comprehensive evaluation value. When the tumor target dose deviation rate is greater than the set threshold and the OAR dose ratio is less than the set threshold, the dose adjustment factor is calculated based on the tumor target volume change rate, and the MLC leaf fine-tuning parameters are generated based on the tumor target displacement vector. When any OAR dose ratio is greater than or equal to the set threshold, the field angle is solved using a constrained optimization algorithm.

[0053] Step 5: Transmit the generated calibration parameters to the linear accelerator, perform field pre-verification based on EPID real-time dose data, and determine whether to perform calibration based on the gamma pass rate;

[0054] Step 6: Store the data from the calibration process into the database. When the data volume reaches the preset amount, perform incremental training on the LSTM neural network to update the tumor regression-dose correction mapping model and the risk organ dose prediction model.

[0055] In one embodiment of the present invention, in step 1, a 4D-CT scanner integrated with the radiotherapy equipment is used to continuously capture the three-dimensional motion trajectory of the tumor target area and adjacent organs at risk during the patient's respiratory cycle at a frequency of 50ms / frame-60ms / frame. The MRI guidance system is synchronously triggered to acquire T1-weighted images every 2min-3min to capture soft tissue deformation. An electronic portal imaging device is used to collect two-dimensional dose distribution data within the portal in real time at a frequency of 10Hz-20Hz, with an accuracy controlled within ±0.1Gy / mm 2 Before treatment, FDG metabolic activity data of tumors are obtained through PET / CT. All data are transmitted to the calibration control module of the treatment planning system in real time to form a four-dimensional dynamic data set including anatomical structure, functional metabolism, motion trajectory and real-time dose, which is recorded as D = {D CT ,D MRI ,D EPID ,D PET}, where D CT Contains the three-dimensional coordinates of the tumor target area and organs in each frame, D EPID is the real-time dose matrix.

[0056] In one embodiment of the present invention: In step 2, for the acquired 4D-CT-D CT With MRI-D MRI Data is elastically registered using an improved spatiotemporal joint registration algorithm. The spatiotemporal joint registration algorithm uses an elastic registration model based on B-spline, combines the time series weight factor to optimize the displacement field, and introduces the time series weight factor λ t And the spatial deformation penalty term Ω(μ) is used to suppress respiratory motion artifacts. The time series weight factor expression is as follows:

[0057] λ t =0.9 (Δt / 300)

[0058] Where Δt is the time interval between the current frame and the initial frame, and the registration energy function is defined as:

[0059]

[0060] Among them, E reg represents the registration energy function, E simis the similarity measure of the grayscale mutual information between 4D-CT and MRI, μ t is the displacement field at time t, α = 0.6, β = 0.3, γ = 0.1 are weight coefficients, and T represents the total number of frames in the time series;

[0061] After registration, the tumor target volume change rate V is extracted from the four-dimensional dynamic data set D rate , Real-time displacement vector of organs at risk and EPID real-time dose hotspot coordinates (x i ,y i ,z i ,D i ), D i represents the actual measured dose value of the dose hotspot area, let D i >1.1D plan For the over-punctual point, D plan It represents the planned dose preset in the treatment plan. The formula for calculating the tumor target volume change rate is as follows:

[0062]

[0063] Among them, V initial is the volume of the first treatment tumor target collected in step 1, V current is the tumor target volume during current treatment.

[0064] In one embodiment of the present invention: In step 3, the displacement vector d of step 2 is used org Perform real-time correction on the initial coordinates (x0, y0, z0) of the organ at risk to obtain dynamic anatomical coordinates The EPID dose matrix D in step 1 is converted into EPID Map to the corrected organ area, calculate voxel dose values ​​and extract peak dose The calculation formula is as follows:

[0065]

[0066] Definition of organ-at-risk dose ratio η org , η org The calculation formula is as follows:

[0067]

[0068] Where v represents the actual region of the organ at risk (OAR) real For each voxel in the image, D(v) represents the actual region of the organ at risk (OAR). real The actual dose value of the voxel v, Preset tolerance threshold for clinical practice and calculate the actual 95% dose coverage value for the tumor target area Combined with planned dose D plan Calculate the dose deviation rate δ D , δ D The calculation is defined as follows:

[0069]

[0070] Introducing the LSTM neural network model, with V rate 、 and historical calibration parameters to predict tumor regression correction factor f(V rate ), construct a comprehensive evaluation function to calculate the evaluation result S, the comprehensive evaluation function is as follows:

[0071] S=0.6δ D +0.3η org +0.1|f(V rate )|

[0072] When S≤0.15, the subsequent calibration process is not triggered, otherwise the subsequent calibration process is triggered.

[0073] In one embodiment of the present invention, in step 4, according to the evaluation value S of step 3 and the deviation type (inadequate dose to the tumor target or excessive dose to the organ at risk), a dual-mode calibration strategy is executed:

[0074] When δ D >10% and all η org <0.8 activates the tumor target dose compensation mode: based on V rate Calculate the dose adjustment factor k using the following formula:

[0075] k=1+0.6×|V rate |

[0076] The value range of k is 1.05≤k≤1.20. When the k value exceeds 1.20, it is 1.20. When the k value is less than 1.05, it is 1.05. The displacement vector of the tumor target area is used as the displacement vector. Generate MLC blade position adjustment parameter Δx j and Δy j , Δx j and Δy j The calculation is as follows:

[0077]

[0078] in, Field plane unit vector;

[0079] When all η org Activate the organ protection mode when the risk is ≥0.8: solve the radiation field angle θ by the constrained gradient descent algorithmnew , define the objective function F as follows:

[0080]

[0081] Among them, F represents the minimized function value, D tol (v) represents the actual region of the organ at risk (OAR) real The clinical tolerance dose threshold of voxel v is constrained by |θ new -θ old |≤15° and the field path avoids the high-density risk area identified by step 2, θ old Indicates the initial angle of the current irradiation field.

[0082] In one embodiment of the present invention, in step 5, the calibration parameters generated in step 4 are transmitted to the linear accelerator, and the calibration parameters include the dose factor k, the blade coordinates (x j +Δx j ,y j +Δy j ) and field angle θ new , based on the EPID data of step one, the field verification is performed, the dose distribution of the verification field is collected, and the γ pass rate of the planned dose after calibration is calculated. When γ ≥ 95%, calibration is performed. When γ < 95%, the δ of step three is used. D Start the backup mechanism: δ D When the value is less than 5%, the historical parameters in the calibration database are called. D When the error is ≥5%, a physicist review reminder will be triggered to ensure the safety and effectiveness of the calibration parameters.

[0083] In one embodiment of the present invention, in step 6, after the calibration is completed, the original data of step 1, the registration parameters of step 2, the evaluation results of step 3, the calibration parameters of step 4 and the verification results of step 5 are stored in the calibration database. When the data accumulates to 50-60 cases, V is automatically extracted. rate With f(V rate ) as training samples, and perform incremental training on the LSTM model in step three, update the tumor regression-dose correction mapping relationship and the risk organ dose prediction model, and form a full closed-loop system of "data acquisition-registration processing-bias assessment-strategy decision-execution verification-model optimization", so as to continuously improve the adaptability of dose calibration to the dynamic changes of individual patients.

[0084] Example 1, please refer to the attached Figure 1 -Attached Figure 2 , dynamic dose calibration for lung cancer patients

[0085] 1. Patient Information and Equipment Configuration

[0086] Patient: A 62-year-old male presented with adenocarcinoma in the right upper lobe of the lung. The tumor was 3.5 cm in maximum diameter and had adjacent organs at risk, including the right hilar lymph nodes, spinal cord, and heart.

[0087] Equipment: Varian TrueBeam linear accelerator (integrated 4D-CT, EPID), Siemens 3T MRI, GEPET / CT.

[0088] Data collection parameters:

[0089] 4D-CT: 50ms / frame, covering the entire respiratory cycle (about 5s), 1mm slice thickness, acquisition frequency 10 frames / respiratory cycle;

[0090] MRI: T1-weighted imaging, 2 min / time, slice thickness 0.8 mm, FOV 350 mm × 350 mm;

[0091] EPID: 10Hz real-time dose acquisition, accuracy ±0.1Gy / mm 2 , pixel resolution 0.5mm×0.5mm;

[0092] PET / CT: 18F-FDG was injected 30 minutes before treatment, and the SUV threshold was set at 2.5 according to clinical guidelines.

[0093] 2. Specific implementation steps

[0094] Step 1: Multimodal dynamic data collection

[0095] 4D-CT scan: The patient lies supine with an abdominal pressure belt fixed. The scan is triggered by respiratory gating, acquiring 300 frames of dynamic chest images per time. The three-dimensional coordinates (x, y, z) of the right lung tumor and the spinal cord displacement trajectory (accuracy ±0.3 mm) are extracted.

[0096] MRI guidance: Synchronously trigger MRI to acquire T1-weighted images, focusing on the hilum and mediastinum, and align with 4D-CT timestamps through the DICOMRT interface;

[0097] EPID real-time monitoring: Each irradiation field triggers EPID acquisition, generates a 512×512 pixel dose matrix, and annotates dose hot spots (D i >1.1D plan area);

[0098] PET / CT metabolic data: extracting tumor SUV max =4.2, calculate the initial tumor target volume V initial =28.5cm 3 .

[0099] Output: 4D dynamic dataset D = {D CT ,D MRI ,DEPID ,D PET}, where D CT Contains the coordinates of the tumor in each frame (such as the displacement vector d of the 100th frame org =(1.2mm,-0.8mm,0.5mm));

[0100] Step 2: Spatiotemporal Joint Registration and Feature Extraction

[0101] Elastic Registration:

[0102] Time series weight factor λ t =0.9 (Δt / 300) , Δt is the interval between the current frame and the first frame (e.g., the 200th frame Δt = 10s, λ t =0.9 (10 / 300) ≈0.996);

[0103] Registration energy function:

[0104] E reg =αE sim (D CT ,D MRI )+βΩ(μ)+γ

[0105] After 50 iterations, the displacement field accuracy is ±0.2 mm.

[0106] Feature extraction:

[0107] Tumor target volume change rate:

[0108]

[0109] Real-time spinal cord displacement vector (Head displacement due to breathing);

[0110] EPID dose hotspot coordinates (15cm, 8cm, 10cm, D i =2.2Gy)(planned dose D plan =2.0Gy).

[0111] Step 3: Dynamic Deviation Quantification and Evaluation

[0112] Organs at risk modifier:

[0113] Spinal cord dynamic coordinates = (x0+0.5mm, y0+1.0mm, z0+0.3mm), EPID dose is mapped by trilinear interpolation, and the (Tolerance Threshold times = 1.8 Gy / time,

[0114] Tumor target dose calculation:

[0115] Actual 95% dose coverage value Dose deviation rate δ D =|2.0-1.9| / 2.0×100%=5%;

[0116] LSTM model prediction:

[0117] Input V rate =-11.2%, and historical calibration parameters, outputting the tumor shrinkage correction factor f(V rate )=0.92;

[0118] Comprehensive evaluation value:

[0119] Trigger calibration;

[0120] Step 4: Dual-mode calibration execution

[0121] Tumor target dose compensation mode (due to δ D =5%<10%, inactive);

[0122] Organ-at-risk protection mode (η org =1.0≥0.8, activated):

[0123] Objective function;

[0124] Constraint |θ new -θ old |≤15°, avoiding the spinal cord path;

[0125] Original field angle θ old =270°, after optimization θ new =255°, calculated γ pass rate = 97% (3mm / 3% standard);

[0126] Step 5: Field Verification and Execution

[0127] EPID verified the field dose distribution, with a γ pass rate of 97% or ≥ 95%. Calibration was performed: the field angle was adjusted to 255°, and the MLC leaf position remained unchanged.

[0128] Step 6: Data closed-loop feedback

[0129] Store this data (V rate =-11.2%,θ new =255° and γ = 97%) to the database, and automatically train the LSTM model after accumulating 50 cases, and update the spinal cord dose prediction accuracy to R 2 =0.94.

[0130] Example 2, please refer to the attached Figure 1 -Attached Figure 2 ,Personalized dose calibration for prostate cancer patients 1. Patient information and equipment configuration

[0131] Patient: A 58-year-old male with prostate cancer, Gleason score 7, and organs at risk including the rectum, bladder, and femoral head.

[0132] Equipment: linear accelerator (integrated 4D-CT, EPID), Philips 1.5T MRI, Siemens PET / CT.

[0133] Data collection parameters:

[0134] 4D-CT: 60ms / frame, covering the entire respiratory cycle (about 6s), slice thickness 0.8mm;

[0135] MRI: T2-weighted imaging, 3 min / time, slice thickness 1 mm, FOV 400 mm × 400 mm;

[0136] EPID: 20Hz real-time dose acquisition, accuracy ±0.1Gy / mm 2 ;

[0137] PET / CT: Assessment of prostate SUV before treatment max =3.8, initial tumor target volume V initial =45.2cm 3 .

[0138] 2. Specific implementation steps

[0139] Step 1: Multimodal dynamic data collection

[0140] 4D-CT scan: The patient lies prone with a moderately full bladder. 250 frames / time are acquired to extract the three-dimensional coordinates of the prostate and the bladder displacement trajectory. Bladder contraction after urination causes the prostate to move upward, and the displacement vector

[0141] MRI guidance: T2-weighted images of the prostate and rectum were obtained, and 4D-CT images were aligned by time stamp to identify rectal wall displacement (accuracy ±0.5 mm) with a resolution of 256 × 256.

[0142] EPID monitoring: Collect the sub-field dose matrix and mark the dose hotspot in the rectal area (D i =1.9Gy, D plan =1.8Gy).

[0143] Step 2: Spatiotemporal Joint Registration and Feature Extraction

[0144] Elastic Registration:

[0145] Time series weight factor λ t =0.9 (Δt / 300) , the interval between the 150th frame and the first frame is Δt = 15s, and we can calculate:

[0146] λ t =0.9 (15 / 300) =0.9 0.05 ≈0.995;

[0147] Rectal displacement vector after registration Prostate volume change rate: V rate =(42.1cm 3 -45.2cm 3 ) / 45.2=-6.9%

[0148] Step 3: Dynamic Deviation Quantification and Evaluation

[0149] Organs at risk modifier:

[0150] After the rectal dynamic coordinates are corrected, the Tolerance threshold =60Gy / 30 times = 2.0Gy / time), dose ratio:

[0151] Tumor target dose calculation:

[0152] Dose deviation rate:

[0153] LSTM model prediction:

[0154] Input V rate =-6.9%, Output f(V rate )=0.95;

[0155] Comprehensive evaluation value: S = 0.6δ D +0.3η org +0.1|f(V rate )|

[0156] =0.6×2.78%+0.3×0.95+0.1×0.95≈0.0167+0.285+0.095

[0157] =0.3967>0.15

[0158] , triggering the calibration process.

[0159] Step 4: Dual-mode calibration execution

[0160] Tumor target dose compensation mode (δ D =2.78%<10%, inactive);

[0161] Organ-at-risk protection mode (η org =0.95≥0.8, activated):

[0162] The objective function F minimizes the rectal dose, constraining the field angle adjustment to be less than or equal to 15°, and the original angle θ old =180°, optimized to θ new =165°(adjusted 15°);

[0163] Dose adjustment factor k = 1 + 0.6 | V rate = 1 + 0.6 × 0.069 = 1.041 (within the range of 1.05-1.20), and the fractionated dose is adjusted to 1.8 × 1.041 = 1.874 Gy;

[0164] Step 5: Field Verification and Execution

[0165] EPID Verificationθ new Pass rate = 96%, calibration performed: adjusted the field angle to 165°, dose adjustment factor k = 1.041, and fine-tuned the MLC blades 0.7 mm toward the head.

[0166] Step 6: Data closed-loop feedback

[0167] The data were stored in the database. After accumulating 60 cases, the LSTM model was trained to update the prostate shrinkage-dose correction relationship, reducing the average dose deviation rate of similar patients to below 3%, and the model prediction error was ≤2%.

[0168] Based on the above two sets of embodiments, it can be concluded that through the real-time synchronous acquisition of multimodal dynamic data from 4D-CT, MRI, EPID, and PET / CT, the dynamic characteristics of the tumor target and organs are accurately extracted by combining a spatiotemporal joint registration algorithm with a time series weight factor. The LSTM neural network is used to construct a comprehensive evaluation function to intelligently judge the calibration conditions. The dose, blade position, and field angle are then dynamically adjusted through a dual-mode calibration strategy. Finally, with the help of data closed-loop feedback, the model is iteratively optimized, effectively correcting the dose deviation caused by tumor shrinkage and organ motion, while ensuring 95% dose coverage of the tumor target and reducing the dose to endangered organs. In both sets of embodiments, this method demonstrates the ability to accurately respond to dynamic changes in anatomical structures in different dynamic change scenarios, such as respiratory motion of lung cancer and organ deformation of prostate cancer. Through the coordinated cooperation of multimodal data interaction, intelligent algorithm decision-making, and model iteration, the radiotherapy dose calibration is transformed from static fixed parameters to dynamic adaptation, significantly improving individualized accuracy and safety, and has outstanding technical innovation and clinical application value.

[0169] Although the present invention is disclosed above with reference to preferred embodiments, this is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent variations, and modifications made to the above embodiments in accordance with the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for calibrating tumor radiotherapy dose, characterized in that: The calibration method comprises the following specific steps: Step 1: Use 4D-CT, MRI, EPID, and PET / CT to synchronously collect the three-dimensional motion trajectory of the patient's tumor target area and organs at risk, soft tissue deformation, two-dimensional dose distribution within the radiation field, and tumor FDG metabolic activity data in real time to form a four-dimensional dynamic data set; Step 2: Use the spatiotemporal joint registration algorithm to perform elastic registration of 4D-CT and MRI data, and introduce the time series weight factor λ t and the spatial deformation penalty term Ω(μ), the registered three-dimensional structure data is obtained, and the volume change rate of the tumor target area, the real-time displacement vector of the organ at risk and the dose hotspot coordinates are extracted; Step 3: Using the real-time displacement vector of the organ at risk to correct the initial coordinates of the organ at risk, mapping the two-dimensional dose distribution data within the radiation field to the corrected organ region, calculating the ratio of the peak dose of the organ at risk to the clinical preset tolerance threshold, calculating the dose deviation rate between the actual 95% dose coverage value of the tumor target area and the preset planned dose, introducing a neural network model to predict the tumor regression correction factor, constructing a comprehensive evaluation function, and calculating the comprehensive evaluation value based on the ratio, dose deviation rate, and tumor regression correction factor; Step 4: Perform dual-mode calibration based on the comprehensive evaluation value. When the tumor target dose deviation rate is greater than the set threshold and the OAR dose ratio is less than the set threshold, the dose adjustment factor is calculated based on the tumor target volume change rate, and the MLC leaf fine-tuning parameters are generated based on the tumor target displacement vector. When any OAR dose ratio is greater than or equal to the set threshold, the field angle is solved using a constrained optimization algorithm. Step 5: Transmit the generated calibration parameters to the linear accelerator, perform field pre-verification based on EPID real-time dose data, and determine whether to perform calibration based on the gamma pass rate; Step 6: Store the data from the calibration process into the database. When the data volume reaches the preset amount, perform incremental training on the LSTM neural network to update the tumor regression-dose correction mapping model and the risk organ dose prediction model.

2. The method for calibrating tumor radiation therapy dose according to claim 1, wherein: In step 1, a 4D-CT scanner integrated with the radiotherapy equipment is used to continuously capture the three-dimensional motion trajectory of the tumor target area and adjacent organs at risk during the patient's respiratory cycle at a frequency of 50ms / frame-60ms / frame. The MRI guidance system is synchronously triggered to acquire T1-weighted images every 2min-3min to capture soft tissue deformation. An electronic portal imaging device is used to collect two-dimensional dose distribution data within the portal in real time at a frequency of 10Hz-20Hz, with an accuracy controlled within ±0.1Gy / mm 2 Before treatment, FDG metabolic activity data of tumors are obtained through PET / CT. All data are transmitted to the calibration control module of the treatment planning system in real time to form a four-dimensional dynamic data set including anatomical structure, functional metabolism, motion trajectory and real-time dose, which is recorded as D = {D CT ,D MRI ,D EPID ,D PET }, where D CT Contains the three-dimensional coordinates of the tumor target area and organs in each frame, D EPID is the real-time dose matrix.

3. The method for calibrating tumor radiation therapy dose according to claim 2, wherein: In the step 2, the collected 4D-CT-D CT With MRI-D MRI Data, the improved spatiotemporal joint registration algorithm is used for elastic registration, and the time series weight factor λ is introduced t And the spatial deformation penalty term Ω(μ) is used to suppress respiratory motion artifacts. The time series weight factor expression is as follows: l t =0.9 (Δt / 300) Where Δt is the time interval between the current frame and the initial frame, and the registration energy function is defined as: Among them, E sim is the similarity measure of the grayscale mutual information between 4D-CT and MRI, μ t is the displacement field at time t, α = 0.6, β = 0.3, γ = 0.1 are weight coefficients, T represents the total number of frames in the time series, and Ω (μ) is the sum of squares of the second-order derivatives of the displacement field; After registration, the tumor target volume change rate V is extracted from the four-dimensional dynamic data set D rate , Real-time displacement vector of organs at risk and EPID real-time dose hotspot coordinates (x i ,y i ,z i ,D i ), D i represents the actual measured dose value of the dose hotspot area, let D i >1.1D plan For the over-punctual point, D plan It represents the planned dose preset in the treatment plan. The formula for calculating the tumor target volume change rate is as follows: Among them, V initial is the volume of the first treatment tumor target collected in step 1, V current is the tumor target volume during current treatment.

4. The method for calibrating tumor radiation therapy dose according to claim 3, wherein: In step three, the displacement vector of step two is used Perform real-time correction on the initial coordinates (x0, y0, z0) of the organ at risk to obtain dynamic anatomical coordinates The EPID dose matrix D in step 1 is converted into EPID Map to the corrected organ area, calculate voxel dose values ​​and extract peak dose The calculation formula is as follows: Definition of organ-at-risk dose ratio η org , η org The calculation formula is as follows: Where v represents the actual region of the organ at risk (OAR) real For each voxel in the image, D(v) represents the actual region of the organ at risk (OAR). real The actual dose value of the voxel v, Preset tolerance threshold for clinical practice and calculate the actual 95% dose coverage value for the tumor target area Combined with planned dose D plan Calculate the dose deviation rate δ D , δ D The calculation is defined as follows: The LSTM neural network model is introduced. The LSTM model contains 3 hidden layers, 128 neurons in each layer, and the input features include the tumor target volume change rate, displacement vector and historical calibration parameters. rate 、 and historical calibration parameters to predict tumor regression correction factor f(V rate ), construct a comprehensive evaluation function to calculate the evaluation result S, the comprehensive evaluation function is as follows: S=0.6δ D +0.3η org +0.1|f(V rate )| When S≤0.15, the subsequent calibration process is not triggered, otherwise the subsequent calibration process is triggered.

5. The method for calibrating tumor radiation therapy dose according to claim 4, wherein: In step 4, the dual-mode calibration strategy is executed according to the evaluation value S and the deviation type in step 3: When δ D >10% and all η org When the dose compensation mode of the tumor target is activated, the dose compensation mode of the tumor target is activated based on V rate Calculate the dose adjustment factor k using the following formula: k=1+0.6×|V rate | The value range of k is 1.05≤k≤1.

20. When the k value exceeds 1.20, it is 1.

20. When the k value is less than 1.05, it is 1.

05. The displacement vector of the tumor target area is used as the displacement vector. Generate MLC blade position adjustment parameter Δx j and Δy j , Δx j and Δy j The calculation is as follows: in, Field plane unit vector; When all η org When the value is ≥0.8, the organ protection mode is activated and the radiation field angle θ is solved by the constrained gradient descent algorithm. new , define the objective function F as follows: Among them, D tol (v) represents the actual region of the organ at risk (OAR) real The clinical tolerance dose threshold of voxel v is constrained by |θ new -θ old |≤15° and the field path avoids the high-density risk area identified by step 2, θ old Indicates the initial angle of the current irradiation field.

6. The method for calibrating tumor radiation therapy dose according to claim 5, characterized in that: In step 5, the calibration parameters generated in step 4 are transmitted to the linear accelerator. The calibration parameters include the dose factor k, the blade coordinates (x j +Δx j ,y j +Δy j ) and field angle θ new , based on the EPID data of step one, the field verification is performed, the dose distribution of the verification field is collected, and the γ pass rate of the planned dose after calibration is calculated. When γ ≥ 95%, calibration is performed, and then according to the δ of step three D Start the backup mechanism, δ D When the value is less than 5%, the historical parameters in the calibration database are called. D When ≥5%, it triggers a reminder for the physiotherapist to review.

7. The method for calibrating tumor radiation therapy dose according to claim 6, wherein: In step 6, after the calibration is completed, the original data of step 1, the registration parameters of step 2, the evaluation results of step 3, the calibration parameters of step 4 and the verification results of step 5 are stored in the calibration database. When the data accumulates to 50-60 cases, V rate With f(V rate ) as training samples, perform incremental training on the LSTM model in step 3, and update the tumor regression-dose correction mapping relationship and the risk organ dose prediction model.

Images