Method and system for calculating optimal dose of TCP and multi-NTCP combination based on deep learning

Abstract

The invention discloses a TCP and multi-NTCP combination optimal dose calculation method and system based on deep learning, relates to the technical field of radiotherapy, and aims to solve the problems of insufficient multi-source data integration, unbalanced multi-organ protection and low individualized precision in traditional dose optimization. The method comprises the following steps: acquiring a CT image of a target patient, a tumor and normal tissue dose-volume histogram (DVH), baseline clinical data and follow-up data after radiotherapy of a patient group; constructing a multi-modal feature data set and an annotation data set through spatial alignment and normalization processing, training a deep learning joint prediction model, synchronously outputting TCP values corresponding to candidate doses and NTCP values of at least three organs at risk, and constructing a fitting curve; and through a single-organ net income formula, a multi-organ comprehensive net income formula and an optimal dose screening formula, quantifying income-risk balance and screening an optimal dose for maximizing the comprehensive net income. According to the invention, through cooperation of deep learning and an exclusive formula system, dynamic quantitative balance of multi-organ unbiased protection and tumor control is realized, individualized precision and clinical landing of radiotherapy dose planning are improved, and the method is suitable for precise radiotherapy scenes of various solid tumors.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy technology, and more particularly to individualized dose planning and efficacy-risk quantification assessment technology in precision radiotherapy. Specifically, it is an optimal dose calculation method and system based on a deep learning-based combination of TCP (probability of tumor control) and multiple NTCP (probability of complications in normal tissues). Background Art

[0002] Radiotherapy is one of the core radical treatments for malignant tumors. Its core objective is to maximize the tumor target dose (TCP) by ensuring the tumor receives a sufficient lethal dose through precise dose planning, while strictly limiting the radiation dose to surrounding organs at risk, such as the lungs, spinal cord, and esophagus, to minimize the non-TCP, ultimately achieving a clinical balance of "maximizing tumor control and minimizing damage to normal tissues." With the popularization of precision radiotherapy technologies such as intensity-modulated radiotherapy (IMRT) and proton therapy, individual differences in radiotherapy dose distribution are becoming increasingly prominent. Traditional dose calculation methods can no longer meet the precise clinical needs for "multi-organ synergistic protection and individualized efficacy optimization," becoming a key technical bottleneck restricting the improvement of radiotherapy efficacy and safety.

[0003] Current radiotherapy dose calculation and optimization techniques have several significant limitations: They lack multi-organ risk assessment (NTCP), traditional methods assess NTCP only for a single organ or simply superimpose multi-organ risks, lacking a unified quantitative model to integrate multi-organ risk, which easily leads to an imbalance in organ protection priorities and fails to achieve balanced protection for multiple organs at risk. Predictive models have poor adaptability; traditional models struggle to characterize the nonlinear relationship and spatial heterogeneity between dose and TCP / NTCP, resulting in insufficient individual adaptability to different patients, limited generalization ability, and predictive accuracy that fails to meet clinical needs. Dose optimization lacks a quantitative system, focusing primarily on single-objective optimization and lacking a comprehensive net benefit assessment system that considers "tumor control benefits - multi-organ risks." The balance between efficacy and risk lacks scientific quantitative basis and relies on physicians' subjective judgment.

[0004] In summary, current radiotherapy dose calculation technology has significant shortcomings in areas such as deep fusion of multimodal data, multi-NTCP collaborative quantification, individualized prediction model construction, and comprehensive net benefit optimization. There is an urgent need to develop an optimal dose calculation method that integrates deep learning and a dedicated quantification formula system. This method should achieve precise individualized dose planning for radiotherapy through multi-organ risk collaborative prediction and benefit-risk quantification balance, providing scientific and efficient dose decision support for clinical practice. This has become a pressing technical problem to be solved in this field. Summary of the Invention

[0005] To address the core technical challenges of existing TCP / NTCP models, including insufficient multi-source heterogeneous data integration capabilities, poor accuracy in capturing dose-space features, weak generalization ability due to limited sample size and label imbalance, and lack of clinical interpretability, this invention proposes a deep learning-based method and system for calculating optimal dose using a combination of TCP and multiple NTCP models. This method and system leverage multimodal data fusion technology and a dedicated deep learning architecture to effectively improve the model's prediction accuracy and generalization ability. Ultimately, it achieves synergistic optimization of maximizing the probability of tumor control (TCP) and minimizing the probability of complications in normal tissues (NTCP) within the tolerance range of normal tissues, providing reliable technical support for individualized radiotherapy dose planning.

[0006] To address the aforementioned problems in the prior art, this invention provides a method for calculating the optimal dose based on a deep learning-based combination of TCP and multiple NTCP protocols, comprising the following steps:

[0007] Collect multi-source clinical data, including CT imaging data of target patients, dose-volume histogram (DVH) data of tumors and normal tissues, and baseline clinical data;

[0008] The data is spatially aligned and normalized to generate a multimodal feature dataset with uniform dimensions.

[0009] Collect follow-up data after radiotherapy of the target patient group, construct a labeled dataset, input the labeled dataset into the deep learning model for joint prediction of TCP and multiple NTCP, and obtain the trained joint prediction model of TCP and multiple NTCP.

[0010] The standardized multimodal feature dataset is input into the trained deep learning model for joint prediction of TCP and multiple NTCP, and the TCP prediction value and NTCP prediction value of at least 3 normal tissues are output for different initial dose points. The prediction fitting curves of TCP and each NTCP are constructed.

[0011] Based on the single-organ net benefit formula, the net benefit of a single normal tissue under a single dosage regimen is calculated; the single-organ net benefit formula is as follows:

[0012] ;in For dosage The TCP value below, For dosage Next NTCP values ​​of normal tissues;

[0013] A single dosing regimen (here) Sum the net gains of all single organs and simplify mathematically:

[0014] (Form 1)

[0015] (Form 2)

[0016]

[0017] (Z-type)

[0018] Summing equation 1 into equation Z and taking the sum, we get...

[0019] Mathematical simplification yields

[0020] This yields the formula for the comprehensive net benefit of multiple organs corresponding to this dosage regimen:

[0021] ;in Indicates the number of normal tissue types that need protection. Indicates the dosage of the i-th type of normal tissue. Net income per organ under the following conditions Indicates dosage The TCP value below, Indicates dosage Down The total NTCP value of normal tissues; extending the multi-organ comprehensive net benefit formula to all candidate dosing regimens to obtain the comprehensive net benefit corresponding to each dosing regimen; for a single candidate dosing regimen Its comprehensive net income formula is:

[0022] ;in Indicates the first Normal tissue in dosage regimen The net benefit per organ; the optimal radiotherapy dose D is determined using the optimal dose screening formula; the optimal dose screening formula is: That is, the combined net benefit from all candidate dosing regimens. Among them, select the one with the largest value. The corresponding radiotherapy dosage regimen is the optimal radiotherapy dose for the target patient.

[0023] Furthermore, the follow-up data includes tumor status assessment results and normal tissue function test data at 1 month, 3 months, 6 months, 12 months, and 24 months after radiotherapy;

[0024] The labeling rules are as follows: if the tumor is stable or shrinking, the TCP label is set to 1; if the tumor is progressing, the TCP label is set to 0; if a corresponding normal tissue complication occurs, the NTCP label of that tissue is set to 1; if no complication occurs, it is set to 0.

[0025] Furthermore, a deep learning-based TCP-NTCP joint prediction model is constructed, including the following steps:

[0026] The input layer is adapted to a multimodal feature dataset. A shared feature encoding network with 3DCNN and fully connected layers is constructed to extract core features. The branch design includes one TCP prediction head and Z NTCP prediction heads. The output layer is activated with Sigmoid and the loss function is binary cross-entropy. The Adam optimization algorithm is used for iterative training until convergence. The model can simultaneously output the TCP and multiple NTCP prediction values ​​corresponding to the candidate dose.

[0027] Furthermore, the formula for net benefit per organ middle, and For the same dose The formula uses the same predicted values ​​from the same source, and both values ​​are in the range of [0,1]. This numerical constraint enables the comparability of net income and ensures that the net income of different normal organizations has a basis for accumulation.

[0028] Optionally, the multi-organ comprehensive net benefit formula In China, through "Z times" "Quantify the overall benefits of tumor control, through..." "Quantify the cumulative risk of multi-organ complications."

[0029] Optionally, the multi-organ comprehensive net benefit formula The core logic is to linearly accumulate the "individual net benefit" of each normal tissue without introducing additional weighting coefficients during the accumulation process. This unbiased accumulation method ensures that all normal tissues that need protection have equal priority in dose assessment, avoiding overemphasis on a single organ.

[0030] Optionally, the optimal dose screening formula The application is limited to: selecting only Candidate dosing regimens participate in the maximization screening if all candidate dosing regimens If the initial dose range needs to be adjusted, the system will output a message to ensure that the screening results meet the clinical baseline requirement that "the benefits outweigh the risks".

[0031] In a preferred embodiment of the present invention, an optimal dose calculation system based on deep learning combining TCP and multiple NTCP is provided, comprising the following modules:

[0032] The data acquisition module is used to collect multi-source clinical data of the target patients and follow-up data of the target patient group after radiotherapy. The multi-source clinical data includes at least CT image data, dose-volume histogram (DVH) data of tumor and normal tissue, and baseline clinical data. The follow-up data includes tumor status assessment results and normal tissue function test data at 1 month, 3 months, 6 months, 12 months and 24 months after radiotherapy.

[0033] The data standardization processing module, which is communicatively connected to the data acquisition module, is used to perform spatial alignment and normalization processing on the multi-source clinical data to generate a multimodal feature dataset with unified dimensions. At the same time, it performs label annotation based on the follow-up data to construct a labeled dataset. The labeling rules are as follows: if the tumor is stable or shrinking, the TCP label is set to 1; if the tumor progresses, the TCP label is set to 0; if a corresponding normal tissue complication occurs, the NTCP label of that tissue is set to 1; if no complication occurs, it is set to 0.

[0034] The deep learning training and prediction module is communicatively connected to the data standardization processing module. It is used to train the model using the labeled dataset, iteratively update the parameters through backpropagation and Adam optimization algorithm until convergence, and obtain the TCP and multiple NTCP joint prediction model. It is also used to receive the multimodal feature dataset, output the TCP prediction value corresponding to different candidate dose points and the NTCP prediction value of at least 3 normal tissues, and construct the prediction fitting curves for TCP and each type of NTCP.

[0035] The single-organ net benefit calculation module, communicatively connected to the deep learning training and prediction module, is used to calculate the net benefit of each normal tissue under a single dosage regimen based on the single-organ net benefit formula, which is: ;

[0036] A multi-organ comprehensive net benefit calculation module, communicatively connected to the single-organ net benefit calculation module, is used to calculate the comprehensive net benefit value corresponding to each candidate dosage regimen based on the multi-organ comprehensive net benefit formula, wherein the multi-organ comprehensive net benefit formula includes: and ;

[0037] The optimal dose screening module, communicatively connected to the multi-organ comprehensive net benefit calculation module, is used to screen the optimal radiotherapy dose D based on the optimal dose screening formula, which is: Select only when filtering Candidate dosing regimens participate in the maximization screening if all candidate dosing regimens If so, the message "Initial dose range needs to be adjusted" will be displayed.

[0038] The output module is communicatively connected to the optimal dose screening module and is used to output the optimal radiotherapy dose and the corresponding TCP prediction value, NTCP prediction value of each normal tissue, net benefit value of a single organ, and comprehensive net benefit value of multiple organs.

[0039] The beneficial effects of this invention include:

[0040] First, multi-NTCP collaborative quantification achieves unbiased protection of multiple organs: by quantifying the benefit-risk difference through the single organ net benefit formula, the comprehensive net benefits of multiple organs are unbiasedly accumulated, avoiding organ priority imbalance, offsetting the cumulative risk of multiple organs, and balancing the protection needs of multiple endangered organs.

[0041] Second, construct a comprehensive net benefit system to quantify the balance between efficacy and risk: Transform tumor control and multi-organ protection into quantitative indicators through a three-level formula system, screen formulas to ensure that "benefits outweigh risks", reduce the subjectivity of physician experience, and provide a scientific basis for decision-making.

[0042] Third, it has strong individualized adaptability and meets the needs of precision radiotherapy: the model takes into account both tumor and organ specificity, and outputs customized dosage plans for individual patient differences, realizing the leap from "group standard" to "individual precision".

[0043] Fourth, the mathematical logic is robust, and the quantitative results are stable: the formula system is progressive and coherent, the parameters are of the same origin, and the unified value range ensures comparability; the "Z times TCP value" design improves clinical suitability. Attached Figure Description

[0044] Figure 1 This application provides a flowchart of the steps for calculating the optimal dose using a combination of TCP and multiple NTCP based on deep learning, as an embodiment of the present application.

[0045] Figure 2 A dose-probability prediction fitting curve of lung cancer TCP and three corresponding normal tissue NTCP based on deep learning provided in an embodiment of the present invention;

[0046] Figure 3 A visualization of the net benefit of the lungs provided in an embodiment of the present invention;

[0047] Figure 4 A visualization of the multi-organ comprehensive net benefit formula provided in this embodiment of the invention;

[0048] Figure 5 This is a visualization of the multi-organ comprehensive net benefit of all candidate doses provided in the embodiments of the present invention;

[0049] Figure 6 A visualization of the optimal radiotherapy dose screening formula based on overall net benefit provided in this embodiment of the invention;

[0050] Figure 7 This is a module architecture diagram of an optimal dose calculation system based on deep learning and combining TCP and multiple NTCP, provided in an embodiment of this application. Detailed Implementation

[0051] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0052] In the following description, specific details, such as particular internal procedures and techniques, are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will appreciate that the invention may be practiced in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of the invention with unnecessary detail.

[0053] The following is in conjunction with the appendix Figure 1-7 The technical solution of the present invention will be further described in detail below.

[0054] Current radiotherapy dose calculation technology has significant technical shortcomings in areas such as multi-source heterogeneous data integration, multi-NTCP collaborative evaluation, and individualized dose optimization. There is an urgent need to develop an optimal dose calculation method based on deep learning-based TCP and multi-NTCP combination.

[0055] In view of the above problems, this application proposes the following embodiments to solve the above technical problems.

[0056] Please see Figure 1 This application provides an optimal dose calculation method based on deep learning for TCP and multiple NTCP combinations, including steps 101-106.

[0057] Step 101: Obtain multi-source clinical data of the target patient, including CT imaging data, dose-volume histogram (DVH) data of tumor and normal tissue, and baseline clinical data of the target patient;

[0058] Step 102: Process the data to generate a multimodal feature dataset with uniform dimensions;

[0059] Step 103: Collect follow-up data after radiotherapy for the target patient group, construct labeled data, use the labeled dataset for training, and obtain the trained TCP and multi-NTCP joint prediction model;

[0060] Step 104: Input the standardized multimodal feature dataset into the trained deep learning model, and output the TCP prediction values ​​corresponding to different initial dose points and the NTCP prediction values ​​of at least 3 normal tissues to predict the fitting curve.

[0061] Step 105: Calculate the net benefit per organ based on the predicted value, calculate the net benefit per organ based on the net benefit per organ, and extend the formula for the comprehensive net benefit per organ to all candidate dosing regimens to obtain the comprehensive net benefit corresponding to each dosing regimen:

[0062] The formula for calculating the net benefit per organ is as follows:

[0063] ;

[0064] in For dosage Below value, For dosage Next normal tissue value;

[0065] By summing and mathematically simplifying the net benefits of all single organs under a single dosage regimen, the formula for the comprehensive net benefit of multiple organs corresponding to that dosage regimen is obtained; the comprehensive net benefit formula for multiple organs is:

[0066] ;

[0067] in Indicates the dosage of the i-th type of normal tissue. Net income per organ under the following conditions Indicates dosage Below value, Indicates dosage Down Total of normal tissues value;

[0068] The multi-organ comprehensive net benefit formula is extended to all candidate dosing regimens to obtain the comprehensive net benefit corresponding to each dosing regimen; for a single candidate dosing regimen... Its comprehensive net income formula is:

[0069] ;

[0070] in Indicates the first Normal tissue in dosage regimen Net income per organ;

[0071] Step 106: Determine the optimal radiotherapy dose D using the optimal dose screening formula;

[0072] The optimal dose selection formula is as follows:

[0073] ;

[0074] That is, the combined net benefit from all candidate dosing regimens. Among them, select the pair with the largest value. The appropriate radiotherapy dosage regimen is determined as the optimal radiotherapy dose for the target patient.

[0075] Step 101 involves collecting multi-source clinical data, including CT image data of the target patient, dose-volume histogram (DVH) data of tumor and normal tissue, and baseline clinical data, specifically including the following:

[0076] Medical imaging data includes high-resolution CT scans used to define the target volume (GTV, CTV, PTV) of the lung cancer tumor and the anatomical structures of surrounding organs at risk (such as the lungs, spinal cord, esophagus, and heart). Dosimetric data includes dose-volume histograms (DVH) of the tumor and normal tissues generated based on the initial radiotherapy plan, extracting key features such as Dmean and Vx (percentage of volume receiving doses of x Gy or higher). Baseline clinical data includes clinical parameters that may affect efficacy and toxicity, such as patient age, sex, tumor stage, pathological type, performance status score, and comorbidities.

[0077] It should be noted that if a patient's data for a certain type is missing (such as missing some DVH features), the mean of the features from patients with the same tumor type can be used for interpolation.

[0078] Step 102 involves spatial alignment and normalization of the data to generate a multimodal feature dataset with uniform dimensions. This process includes the following steps:

[0079] Continuous features (such as dose values, volume parameters, age, etc.) are normalized or standardized to eliminate the influence of dimensions and map them to the range of [0,1] or conforming to a normal distribution. The processed image features, DVH features, and clinical features are then fused to construct a multimodal feature dataset with unified dimensions and structure, which serves as the training set for the model.

[0080] Step 103, which involves constructing the labeled dataset, specifically includes the following steps:

[0081] First, a labeled dataset is constructed: long-term follow-up data of the target patient group (e.g., lung cancer patients in specific locations) after radiotherapy is collected. The follow-up data should include multiple time points, namely imaging assessments at 1, 3, 6, 12, and 24 months after radiotherapy to determine tumor status, and corresponding assessments of toxicity to normal tissues. The labeled dataset is then input into a deep learning model for joint TCP and multiple NTCP prediction for training, resulting in a trained joint TCP and multiple NTCP prediction model.

[0082] It should be noted that the TCP tag is set to 1 for the corresponding dose regimen if the tumor achieves complete remission, partial remission, or stability during the follow-up period, indicating successful control; if the tumor progresses, the tag is set to 0, indicating control failure. The NTCP tag is set to 1 for each pre-defined organ at risk (Z≥3) if grade ≥2 related radiation damage or complications occur during the follow-up period, indicating that complications have occurred; otherwise, it is set to 0, indicating that no complications have occurred.

[0083] Step 103, which involves constructing a deep learning model for joint prediction of TCP and multiple NTCP, specifically includes the following steps:

[0084] To construct a hierarchical network structure, a two-layer architecture of "shared feature encoding - multi-task prediction" is designed based on the characteristics of multimodal feature datasets, enabling deep fusion of multi-source information and synchronous prediction of TCP / multi-NTCP.

[0085] Input layer adaptation: To address the heterogeneity of multimodal feature datasets, CT image data is converted into a three-dimensional tensor (dimensions H×W×D, where H, W, and D represent the image height, width, and depth, respectively). Key dosimetric features such as Dmean, V20, and V30 extracted from DVH data, along with baseline clinical data, are converted into one-dimensional vectors. Through feature concatenation, the three-dimensional image tensor and one-dimensional structured features are fused into a unified-dimensional input tensor (dimensions H×W×D×C, where C is the number of feature channels), ensuring that the input data is adapted to the processing requirements of the 3DCNN network.

[0086] Shared Feature Encoding Network Construction: A shared feature extraction module is constructed using a 3D Convolutional Neural Network (3DCNN) combined with fully connected layers. The 3DCNN part consists of 4-6 convolutional layers with kernel sizes of 3×3×3, 3×3×3, 2×2×2, and 2×2×2, respectively, and a stride of 1 for each layer. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to enhance feature representation and alleviate the gradient vanishing problem. After each convolutional layer, alternating max-pooling layers (2×2×2 kernel size, stride 2) are applied to achieve feature dimensionality compression and key information preservation. The feature tensor processed by the 3DCNN is converted into a one-dimensional feature vector through a flattening operation and then connected to two fully connected layers (1024 and 512 neurons respectively) to further extract high-level abstract features, ultimately outputting a shared core feature vector with a dimension of 256.

[0087] Multi-task prediction head design: Based on a shared core feature vector, one TCP prediction head and Z NTCP prediction heads are designed (Z is the number of normal tissue types to be protected, and Z≥3). Each prediction head consists of a fully connected layer (64 neurons) and an output layer. The output layer uses the Sigmoid activation function to map the prediction results to the [0,1] interval, consistent with the probability range of TCP / NTCP, ensuring that the predicted values ​​have direct clinical interpretability. Specifically, the TCP prediction head outputs a single scalar value representing the probability of tumor control, and each NTCP prediction head outputs a single scalar value representing the probability of complications in the corresponding normal tissue.

[0088] The design and derivation of the loss function are crucial. Since the model involves multi-task joint training (1 TCP prediction task + Z NTCP prediction tasks), a joint loss function needs to be constructed to balance the training weights of each task and ensure the model's adaptability to different tasks.

[0089] Single Task Loss Function: For each prediction task (TCP prediction or NTCP prediction), since the labels are binary (TCP label 0 / 1, NTCP label 0 / 1), Binary Cross Entropy (BCE) is used as the loss function for a single task. For the k-th task (k=0 corresponds to the TCP task, k=1,2,...,Z corresponds to Z NTCP tasks), the loss function formula is:

[0090]

[0091] Where N is the number of training samples, This represents the true label (0 or 1) of the m-th sample in the k-th task. This represents the predicted probability value of the m-th sample in the k-th task by the model. It is the natural logarithm.

[0092] Joint Loss Function Construction: To ensure that each task receives equal attention during training and to avoid a single task dominating the training process, a joint loss function is constructed using an equal-weighted summation method. The formula is as follows:

[0093]

[0094] in, The loss value for the TCP prediction task. Let Z be the sum of the loss values ​​of the Z NTCP prediction tasks. This is a normalization coefficient to ensure that the magnitude of the joint loss value is consistent with the loss value of a single task, which facilitates gradient adjustment during training.

[0095] Model training and convergence optimization, based on the constructed network structure and loss function, employs iterative training to optimize model parameters. The specific process is as follows:

[0096] Initialization parameter settings: Randomly initialize all trainable parameters in the network (convolutional layer weights, fully connected layer weights, bias terms) using the He normal distribution (suitable for parameter initialization of ReLU activation function) to ensure that the mean of the initial parameters is 0 and the variance is adapted to the input and output dimensions of the network layers, laying the foundation for training convergence.

[0097] Forward propagation computation: The multimodal features from the labeled dataset are input into the model, the core features are extracted through a shared feature encoding network, and then the forward propagation computation of each prediction head is performed to obtain the prediction probability value of each sample on the TCP task and Z NTCP tasks. (m=1,2,...,N; k=0,1,...,Z).

[0098] Loss calculation and backpropagation: Based on the joint loss function formula above, substitute the true labels... and predicted probability Calculate the joint loss value for the current iteration round. Based on the backpropagation algorithm, starting from the joint loss value, the partial derivatives (gradients) of the loss function with respect to the parameters of each layer are calculated sequentially. ( (representing any trainable parameter in the network), to obtain the gradient update direction of each parameter.

[0099] Adam optimizer parameter update: The Adam optimization algorithm is used to iteratively update the network parameters. This algorithm combines momentum gradient descent and adaptive learning rate adjustment mechanisms, which can effectively improve training stability and convergence speed. The parameter update formula is as follows:

[0100] Calculate the momentum term:

[0101] Calculate the second momentum term:

[0102] Momentum term deviation correction:

[0103] Second-order momentum term deviation correction:

[0104] Parameter update:

[0105] Where t is the current iteration round, , These are the momentum term and second-order momentum term from the previous cycle, respectively. (Default value is 0.9) (The default value is 0.999) is the momentum decay coefficient. (Initial value 0.001) is the learning rate. (Default value) To prevent the minimum value where the denominator is 0.

[0106] Convergence criterion: Set the total number of training iterations (default 200) and an early stopping mechanism to avoid overfitting. Every 10 training iterations, calculate the joint loss value of the model using the validation set. If the validation set loss value does not decrease for 20 consecutive iterations and the fluctuation range is ≤0.001, the model is considered to have converged, and training stops. If the early stopping mechanism is not triggered, it continues to iterate until the end of the total number of iterations. Finally, save the model parameters in the converged state to obtain the trained TCP and multi-NTCP joint prediction model.

[0107] It should be noted that the deep learning model for joint prediction of TCP and multiple NTCP is a multi-task deep learning model based on a hybrid architecture of 3DCNN and fully connected layers. The model takes a multimodal feature dataset as input and designs a shared feature encoding network, which then branches into multiple task heads: one TCP prediction head, outputting a scalar representing the probability of tumor control (TCP); and Z NTCP prediction heads, each outputting a scalar corresponding to the probability of complications (NTCP) in a specific normal tissue.

[0108] In step 104, the TCP prediction value corresponding to different initial dose points and the NTCP prediction value of at least 3 normal tissues are output, and the prediction fitting curves corresponding to TCP and each type of NTCP are constructed.

[0109] Based on clinical guidelines and the patient's specific circumstances, a reasonable range of candidate radiotherapy doses should be established for exploration.

[0110] For example, in this embodiment for lung cancer, based on the clinical radiotherapy fractionation dose standard (in this embodiment, the single fraction dose is 2 Gy), the continuous range of 20 Gy to 70 Gy is discretized into 26 candidate dose points: =20, 22, 24, ..., 70 Gy. This discrete method ensures that each candidate dose point can directly correspond to the actual number of fractions in radiotherapy (e.g., 50 Gy corresponds to 25 fractions, with a single fraction dose of 2 Gy), and can be applied without additional adjustments.

[0111] For each candidate dose point in the range Adjust the planned dose distribution parameters in the input data so that the model can simultaneously predict the dose at that point. Value and Z Values ​​(i=1,2,...,Z). For each dose point... and its corresponding and each The values ​​are plotted on a dose-probability coordinate system. A logistic function is used for fitting, generating one smooth TCP prediction curve and Z NTCP prediction curves.

[0112] For example, the effectiveness of the fitting function is verified by cross-validation: the original prediction data is randomly divided into 5 subsets, and the fitting is repeated 5 times using a 4-fold training and 1-fold validation method to ensure that the mean square error fluctuation of the fitting of each subset is ≤0.01; in this embodiment, the mean absolute value of the residual of the logistic function fitting is 0.02, and the mean square error fluctuation of the cross-validation is 0.008, which meets the clinical accuracy requirements.

[0113] The formula for calculating the net benefit per organ in step 105 is as follows: .

[0114] in This represents the nth discrete radiotherapy dose regimen; Indicates dosage regimen The probability of tumor control is as follows; Indicates dosage regimen The probability of complications in the Zth type of normal tissue; Indicates dosage regimen Below, the net single-organ benefit of "tumor control benefit" versus "risk of type Z normal tissue"; Indicates dosage regimen Below, the combined net benefit of all Z types of normal tissues; Indicates the optimal radiotherapy dose;

[0115] It should be noted that the larger the difference, the higher the net benefit of the dosage regimen for that organ. Because and The values ​​are all in the range [0,1]. The value range is [-1, 1], which ensures the comparability of net benefits between different organs.

[0116] For example, in the case of a lung cancer patient, three organs at risk requiring priority protection can be selected: the lung, spinal cord, and esophagus. In this case, [the following is used:] Dosage, to obtain lung cancer lungs spinal cord esophagus .

[0117] For example, based on the above formula for net benefit per organ available:

[0118] (Form 1)

[0119] (Form 2)

[0120] (Form 3)

[0121] Adding equation 1 to equation 3 and summing the results, we get...

[0122] .

[0123] Mathematical simplification yields .

[0124] Derivation and calculation formula of comprehensive net benefit of multiple organs: .

[0125] in Indicates the number of normal tissue types that need protection. Indicates the dosage of the i-th type of normal tissue. Net income per organ under the following conditions Indicates dosage Below value, Indicates dosage Down Total of normal tissues value.

[0126] It should be noted that the comprehensive net income =Total Global Gain - Total Cumulative Risk. This formula, through linear summation without introducing artificial weights, assigns equal importance to each organ requiring protection.

[0127] The multi-organ comprehensive net benefit formula is extended to all candidate dosing regimens to obtain the comprehensive net benefit corresponding to each dosing regimen; for a single candidate dosing regimen... Its comprehensive net income formula is:

[0128] ;

[0129] in Indicates the first Normal tissue in dosage regimen Net income per organ.

[0130] In step 106, the optimal radiotherapy dose D is determined using the optimal dose screening formula.

[0131] The optimal dose selection formula is as follows:

[0132]

[0133] That is, the combined net benefit from all candidate dosing regimens. Among them, select the one with the largest value. The corresponding radiotherapy dosage regimen is used as the optimal radiotherapy dose for the target patient.

[0134] It should be noted that the formula for net benefit per organ is... Formula for comprehensive net benefit of multiple organs They form a progressive relationship, with the single-organ net benefit formula as the basic level quantification unit and the multi-organ comprehensive net benefit formula as the integration level quantification unit. Together, they achieve hierarchical quantification "from single organ to multiple organs". Moreover, the mathematical logic of the two formulas is consistent and the parameters are from the same source, ensuring the consistency and accuracy of the quantification results.

[0135] It should be noted that the optimal dose screening formula The output directly corresponds to the clinically feasible radiotherapy dose value, and this dose value must meet the following requirements: Compared to other candidate dosing regimens The difference should be at least 5% higher. This threshold should be used to avoid frequent changes in the dosage regimen due to small numerical differences, thereby improving the stability of clinical applications.

[0136] Please see Figure 2 , Figure 2 A dose-probability prediction fitting curve of lung cancer TCP and three corresponding normal tissue NTCP based on deep learning provided in an embodiment of the present invention;

[0137] First, the single fractionation dose of tumor radiotherapy was determined to be 2 Gy. Based on the deep learning model, the raw TCP data points of lung cancer and three corresponding normal tissues (lung, esophagus, and spinal cord) at doses of 20-70 Gy were predicted.

[0138] Since the dose-probability relationship between TCP and NTCP exhibits a typical S-shaped variation, this embodiment selects the Logistic function as the fitting function, whose general form is:

[0139]

[0140] in: denoted as D, where a is the TCP / NTCP probability value corresponding to dose D; a is the maximum asymptotic value of the curve (between 0 and 1, reflecting the maximum probability); b is the inflection point dose (i.e., the dose corresponding to when the probability reaches a / 2, D50); and c is a parameter related to the slope of the curve (slope sensitivity k=1 / c, the smaller c is, the larger k is, and the more sensitive the probability is to changes in dose).

[0141] against Figure 2 For different indicators, the original data points are fitted using the least squares method to obtain the fitting function and parameters corresponding to each indicator:

[0142] Fitting function for Lung Cancer TCP:

[0143] in: This is the maximum asymptotic value for TCP (approximately 0.96). =0.25 (i.e.) =4); =35Gy;

[0144] Fitting function for lung NTCP:

[0145] in: lung The maximum asymptotic value; =0.15 (i.e.) ≈6.67); =50Gy;

[0146] Fitting function for esophageal NTCP:

[0147] in: For esophagus The maximum asymptotic value; =0.15 (i.e.) ≈6.67); =55Gy;

[0148] Fitting function for spinal cord NTCP:

[0149] in: This represents the maximum asymptotic value of NTCP in the spinal cord. =0.15 (i.e.) ≈6.67); =60Gy;

[0150] Subsequently, by substituting the dose values ​​within the range of 20–70 Gy into the above fitting function, continuous values ​​corresponding to each fitted curve in the fitted value graph of lung cancer TCP and three normal tissue NTCP at each dose can be obtained. Verification showed that the mean square error between the fitting function and the original data points was ≤0.03, and the fitting effect met the clinical accuracy requirements.

[0151] Please see Figure 3 , Figure 3 The visualization of the net benefit of the lung provided in the embodiment of the present invention corresponds to the technical step of "single organ net benefit calculation" in the present invention.

[0152] Combination Figure 2 The resulting fitting function, taking a dose of 36 Gy as an example: Substituting D = 36 Gy into the fitting function, we get... (36)≈0.588、 ≈0.112, the difference between the two is 0.588−0.112=0.476.

[0153] From the curve trend: as the dose increases, the TCP fitting curve rises rapidly first and then tends to flatten, while the lung NTCP fitting curve shows a slow upward trend; therefore, the difference between the two first increases with the increase of dose, reaching a high level in the 42~50Gy range. After exceeding 50Gy, due to the slowdown of TCP growth and the continuous increase of lung NTCP, the net benefit growth gradually tends to flatten.

[0154] Please see Figure 4 , Figure 4 This is a visualization of the multi-organ comprehensive net benefit formula provided in the embodiments of the present invention, corresponding to the technical step of "multi-organ comprehensive net benefit calculation" in the present invention.

[0155] Taking D=36Gy as an example, combined with Figure 2 , Figure 3 The fitted values ​​are then used for calculation:

[0156] Value: 3× =3×0.588≈1.764;

[0157] Values: Lung NTCP (approx. 0.112) + Esophagus NTCP (approx. 0.068) + Spinal cord NTCP (approx. 0.031) ≈ 0.211;

[0158] The overall net return for D=36Gy is approximately 1.764−0.211≈1.553.

[0159] Please see Figure 5 , Figure 5 This is a visualization of the comprehensive net benefit of all candidate doses provided in the embodiments of the present invention, corresponding to the technical step of "comprehensive net benefit assessment of the entire candidate dose range" in the present invention.

[0160] Judging from the trend of the curve:

[0161] In the 20-50 Gy range: the overall net benefit increases rapidly with increasing dose—during this phase, the increase in lung cancer TCP is much greater than the sum of the increases in NTCP of the lung, esophagus, and spinal cord (combined with...). Figure 1 The fitted curves show that TCP rises rapidly in the 20-40 Gy range, while multi-organ NTCP grows more slowly, thus the overall net return continues to improve.

[0162] Around 50~52Gy: The overall net return reaches its peak (approximately 2.25) - at this point, TCP is close to saturation (maximum asymptotic value of approximately 0.96), while the growth rate of the sum of NTCP across multiple organs begins to approach the growth rate of TCP, and the overall net return enters a plateau period and reaches its maximum value.

[0163] After 52 Gy: The overall net benefit gradually decreases with increasing dose - at this stage, TCP basically stops increasing, but the total NTCP of multiple organs continues to rise with increasing dose, and the "benefit-risk" balance between the two is broken, resulting in a decline in the overall net benefit.

[0164] Please see Figure 6 , Figure 6 The image shown is a visualization of the optimal radiotherapy dose screening formula based on comprehensive net benefit provided in the embodiments of the present invention, corresponding to the core technical aspect of the "optimal dose screening" of the present invention.

[0165] Taking 50Gy as an example, substitute the values ​​into the fitting functions for calculation:

[0166] Lung cancer TCP value: ≈0.938 (close to the maximum asymptotic value of TCP 0.96, indicating sufficient tumor control).

[0167] Multi-organ NTCP sum: ≈0.475 + 0.108 + 0.034 = 0.617;

[0168] Overall net income: 3 × 0.938 − 0.617 ≈ 2.814 − 0.617 = 2.197.

[0169] The optimal dose point (50 Gy) is the maximum value of the overall net benefit fitting curve, representing the highest overall net benefit at 50 Gy within the candidate dose range of 20-70 Gy.

[0170] At this point, the overall net benefit is maximized, proving that 50 Gy is the optimal radiotherapy dose for the patient.

[0171] For example, the patient was given a radiotherapy dose of 50 Gy. A 24-month follow-up after radiotherapy showed that the tumor was completely remission (TCP=1), and no ≥ grade 2 complications occurred in the lungs, esophagus, or spinal cord (total NTCP=0.45). The clinical benefit was in line with expectations.

[0172] Please see Figure 7 , Figure 7 This application provides a module architecture diagram of an optimal dose calculation system based on deep learning and combining TCP and multiple NTCP, comprising:

[0173] The data acquisition module is used to collect multi-source clinical data of the target patients and follow-up data of the target patient group after radiotherapy. The multi-source clinical data includes at least CT image data, dose-volume histogram (DVH) data of tumor and normal tissue, and baseline clinical data. The follow-up data includes tumor status assessment results and normal tissue function test data at 1 month, 3 months, 6 months, 12 months and 24 months after radiotherapy.

[0174] The data standardization processing module, which is communicatively connected to the data acquisition module, is used to perform spatial alignment and normalization processing on the multi-source clinical data to generate a multimodal feature dataset with unified dimensions. At the same time, it performs label annotation based on the follow-up data to construct a labeled dataset. The labeling rules are as follows: if the tumor is stable or shrinking, the TCP label is set to 1; if the tumor progresses, the TCP label is set to 0; if a corresponding normal tissue complication occurs, the NTCP label of that tissue is set to 1; if no complication occurs, it is set to 0.

[0175] The deep learning training and prediction module is communicatively connected to the data standardization processing module. It is used to train the model using the labeled dataset, iteratively update the parameters through backpropagation and Adam optimization algorithm until convergence, and obtain the TCP and multiple NTCP joint prediction model. It is also used to receive the multimodal feature dataset, output the TCP prediction value corresponding to different candidate dose points and the NTCP prediction value of at least 3 normal tissues, and construct the prediction fitting curves for TCP and each type of NTCP.

[0176] The single-organ net benefit calculation module, communicatively connected to the deep learning training and prediction module, is used to calculate the net benefit of each normal tissue under a single dosage regimen based on the single-organ net benefit formula, which is: ;

[0177] A multi-organ comprehensive net benefit calculation module, communicatively connected to the single-organ net benefit calculation module, is used to calculate the comprehensive net benefit value corresponding to each candidate dosage regimen based on the multi-organ comprehensive net benefit formula, wherein the multi-organ comprehensive net benefit formula includes: and ;

[0178] The optimal dose screening module, communicatively connected to the multi-organ comprehensive net benefit calculation module, is used to screen the optimal radiotherapy dose D based on the optimal dose screening formula, which is: Select only when filtering Candidate dosing regimens participate in the maximization screening if all candidate dosing regimens If so, the message "Initial dose range needs to be adjusted" will be displayed.

[0179] The output module is communicatively connected to the optimal dose screening module and is used to output the optimal radiotherapy dose and the corresponding TCP prediction value, NTCP prediction value of each normal tissue, net benefit value of a single organ, and comprehensive net benefit value of multiple organs.

[0180] This method integrates multimodal clinical data with a 3DCNN multi-task deep learning model to achieve simultaneous and accurate prediction of TCP and multiple NTCP. Relying on a three-level quantitative formula system, it unbiasedly balances the benefits of tumor control with the risks of multi-organ damage, and selects the individualized optimal dose where the benefits outweigh the risks. It can be implemented clinically without additional adjustments, significantly improving the accuracy and safety of radiotherapy.

[0181] The system constructs a fully automated closed-loop architecture, with each module working together to complete data acquisition, standardization, model training and prediction, net benefit calculation and dose screening; it can simultaneously output multi-dimensional evaluation data, and the results are traceable and verifiable; it has good scalability, is compatible with a variety of radiotherapy technologies, and provides efficient and comprehensive dose decision support for clinical practice.

[0182] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements all the steps of the deep learning-based TCP and multiple NTCP combination optimal dose calculation method described in the above embodiments.

[0183] This application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements all the steps of the deep learning-based TCP and multiple NTCP combination optimal dose calculation method in the above embodiments.

[0184] Although the present application is described with reference to examples, it is for illustrative purposes only and not for limiting the present application. Any changes, additions or deletions to the implementation of the electronic device or storage medium may be made without departing from the scope of protection of the present application.

[0185] It should be noted that the specific embodiments described above are only for explaining the present invention and are not intended to limit the present invention. Those skilled in the art should understand that the technical parameters, model structures, etc. involved in the following embodiments can be adaptively adjusted according to actual application scenarios, and still fall within the protection scope of the present invention.

Claims

1. A method for calculating the optimal dose of TCP and multiple NTCP combinations based on deep learning, characterized in that, include: Collect multi-source clinical data, including CT imaging data of target patients, dose-volume histogram (DVH) data of tumors and normal tissues, and baseline clinical data; The data is spatially aligned and normalized to generate a multimodal feature dataset with uniform dimensions. Collect follow-up data after radiotherapy of the target patient group, construct a labeled dataset, input the labeled dataset into the deep learning model for joint prediction of TCP and multiple NTCP, and obtain the trained joint prediction model of TCP and multiple NTCP. The standardized multimodal feature dataset is input into the trained TCP and NTCP joint prediction model, and the TCP prediction value and NTCP prediction value of at least 3 normal tissues are output for different initial dose points. The prediction fitting curves of TCP and each NTCP are constructed. Based on the single-organ net benefit formula, the net benefit of a single normal tissue under a single dosage regimen is calculated; the single-organ net benefit formula is as follows: ； in For dosage The TCP value below, For dosage The NTCP value of the Zth normal tissue; By summing and mathematically simplifying the net benefits of all single organs under a single dosage regimen, the formula for the comprehensive net benefit of multiple organs corresponding to that dosage regimen is obtained; the comprehensive net benefit formula for multiple organs is: ; in Indicates the dosage of the i-th type of normal tissue. Net income per organ under the following conditions Indicates dosage The TCP value below, Indicates dosage Total NTCP values ​​for the following Z types of normal tissues; The multi-organ comprehensive net benefit formula is extended to all candidate dosing regimens to obtain the comprehensive net benefit corresponding to each dosing regimen; for a single candidate dosing regimen... Its comprehensive net income formula is: ； in Indicates the dosage regimen of the i-th normal tissue. Net income per organ; The optimal radiotherapy dose D is determined using an optimal dose screening formula; the optimal dose screening formula is: ； That is, the combined net benefit from all candidate dosing regimens. Among them, select the one with the largest value. The corresponding radiotherapy dosage regimen is the optimal radiotherapy dose for the target patient.

2. The method for calculating the optimal dose of TCP and multiple NTCP combinations based on deep learning according to claim 1, characterized in that, The follow-up data includes tumor status assessment results and normal tissue function test data at 1 month, 3 months, 6 months, 12 months and 24 months after radiotherapy; the labeling rules are as follows: if the tumor status is stable or shrinking, the TCP label is set to 1; if the tumor progresses, the TCP label is set to 0; if a corresponding normal tissue complication occurs, the NTCP label of that tissue is set to 1; if no complication occurs, it is set to 0.

3. The method for calculating the optimal dose of TCP and multiple NTCP combinations based on deep learning according to claim 1, characterized in that, The construction of the deep learning model for joint prediction of TCP and multiple NTCP is specifically as follows: The input layer is adapted to a multimodal feature dataset. A shared feature encoding network with 3DCNN and fully connected layers is constructed to extract core features. The branch design includes one TCP prediction head and Z NTCP prediction heads. The output layer is activated with Sigmoid and the loss function is binary cross-entropy. The Adam optimization algorithm is used for iterative training until convergence. The model can simultaneously output the TCP and multiple NTCP prediction values ​​corresponding to the candidate dose.

4. The method for calculating the optimal dose of TCP and multiple NTCP combinations based on deep learning according to claim 1, characterized in that, The formula for net benefit per organ middle, and For the same dose The formula uses the same predicted values ​​from the same source, and both values ​​are in the range of [0,1]. This numerical constraint enables the comparability of net income and ensures that the net income of different normal organizations has a basis for accumulation.

5. The method for calculating the optimal dose of TCP and multiple NTCP combinations based on deep learning according to claim 1, characterized in that, The formula for comprehensive net benefit of multiple organs In China, through "Z times" "Quantify the overall benefits of tumor control, through..." "Quantify the cumulative risk of multi-organ complications." 6. The method for calculating the optimal dose of TCP and multiple NTCP combinations based on deep learning according to claim 1, characterized in that, The formula for comprehensive net benefit of multiple organs The core logic is to linearly accumulate the "individual net benefit" of each normal tissue without introducing additional weighting coefficients during the accumulation process. This unbiased accumulation method ensures that all normal tissues that need protection have equal priority in dose assessment, avoiding overemphasis on a single organ.

7. The method for calculating the optimal dose of TCP and multiple NTCP combinations based on deep learning according to claim 1, characterized in that, The optimal dose screening formula The application is limited to: selecting only Candidate dosing regimens participate in the maximization screening if all candidate dosing regimens If the initial dose range needs to be adjusted, the output will be "Initial dose range needs to be adjusted" to ensure that the screening results meet the clinical baseline requirement that "benefits outweigh risks".

8. An optimal dose calculation system based on deep learning combining TCP and multiple NTCP, characterized in that, The system includes: The data acquisition module is used to collect multi-source clinical data of the target patients and follow-up data of the target patient group after radiotherapy. The multi-source clinical data includes at least CT image data, dose-volume histogram (DVH) data of tumor and normal tissue, and baseline clinical data. The follow-up data includes tumor status assessment results and normal tissue function test data at 1 month, 3 months, 6 months, 12 months and 24 months after radiotherapy. The data standardization processing module, which is communicatively connected to the data acquisition module, is used to perform spatial alignment and normalization processing on the multi-source clinical data to generate a multimodal feature dataset with unified dimensions. At the same time, it performs label annotation based on the follow-up data to construct a labeled dataset. The labeling rules are as follows: if the tumor is stable or shrinking, the TCP label is set to 1; if the tumor progresses, the TCP label is set to 0; if a corresponding normal tissue complication occurs, the NTCP label of that tissue is set to 1; if no complication occurs, it is set to 0. The deep learning training and prediction module is communicatively connected to the data standardization processing module. It is used to train a deep learning model for joint prediction of TCP and multiple NTCP using the labeled dataset to obtain a joint prediction model of TCP and multiple NTCP. It is also used to receive the multimodal feature dataset, output the TCP prediction value corresponding to different candidate dose points and the NTCP prediction value of at least three normal tissues, and construct the prediction fitting curves for TCP and each type of NTCP. The single-organ net benefit calculation module, communicatively connected to the deep learning training and prediction module, is used to calculate the net benefit of each normal tissue under a single dosage regimen based on the single-organ net benefit formula, which is: ; A multi-organ comprehensive net benefit calculation module, communicatively connected to the single-organ net benefit calculation module, is used to calculate the comprehensive net benefit value corresponding to each candidate dosage regimen based on the multi-organ comprehensive net benefit formula, wherein the multi-organ comprehensive net benefit formula includes: and ; The optimal dose screening module, communicatively connected to the multi-organ comprehensive net benefit calculation module, is used to screen the optimal radiotherapy dose D based on the optimal dose screening formula, which is: Select only Candidate dosing regimens participate in the maximization screening if all candidate dosing regimens If so, the message "Initial dose range needs to be adjusted" will be output. The output module is communicatively connected to the optimal dose screening module and is used to output the optimal radiotherapy dose and the corresponding TCP prediction value, NTCP prediction value of each normal tissue, net benefit value of a single organ, and comprehensive net benefit value of multiple organs.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for calculating the optimal dose of TCP and multiple NTCP based on deep learning as described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for calculating the optimal dose of TCP and multiple NTCP based on deep learning as described in any one of claims 1-6.

Images