Method and system for generating dosage scheme by TCP and NTCP based on deep learning

Abstract

The invention discloses a deep learning-based TCP and NTCP dosage scheme generation method and system, and the method comprises the following steps: obtaining multi-source heterogeneous data of a target patient group, carrying out the standardization processing of the multi-source heterogeneous data, and obtaining a data set with consistent dimensions; labeling corresponding TCP and NTCP labels according to the follow-up visit result of the patient; inputting the training data set into a TCP-NTCP combined prediction model, learning nonlinear correlation between data and TCP / NTCP, outputting prediction results under different doses, and substituting the prediction results into a TCP-NTCP combined dose-response function to establish a combined dose-response curve; and substituting into an optimal screening constraint condition formula of a radiotherapy dose scheme, and solving an optimal dose value considering both tumor control and normal tissue protection. Through multi-modal data fusion and a deep learning architecture, the model prediction precision and generalization ability are improved, collaborative optimization of TCP maximization and NTCP minimization in a normal tissue tolerance range is realized, and technical support is provided for radiotherapy individualized dose planning.

Description

Technical Field

[0001] This invention relates to the field of medical data processing, and in particular to a method and system for generating dosage schemes using TCP and NTCP based on deep learning. Background Technology

[0002] Radiotherapy is a core treatment for cancer, and its efficacy and safety depend on the precise optimization of the radiotherapy plan: ensuring the tumor-killing dose in the target area to improve the probability of tumor control (TCP) while reducing the radiation dose to normal tissues to reduce the probability of complications (NTCP). TCP and NTCP models are core tools for evaluating and optimizing radiotherapy plans, directly impacting treatment outcomes and patient prognosis. With the development of precision technologies such as intensity-modulated radiotherapy (IMRT) and proton therapy, individual differences in radiotherapy dose distribution have become more prominent. Traditional population-based statistical dose-response models can no longer meet the needs of personalized treatment. Clinically, there is an urgent need for TCP / NTCP models that can accurately capture individual dose characteristics, patient heterogeneity, and the correlation between efficacy and complications. Existing TCP / NTCP models have significant limitations: First, traditional statistical models, such as the LKB model and the Niemierko probability model, rely on population data fitting, assume uniform dose distribution, do not consider the spatial heterogeneity of actual radiotherapy, have poor adaptability to individual differences, and lack prediction accuracy in complex dose scenarios; Second, traditional machine learning models, such as support vector machines and random forests, although they can incorporate some clinical features, have weak ability to extract spatial features from high-resolution data such as three-dimensional dose matrices, their generalization performance is affected by sample size and data bias, and they lack clinical interpretability, making them difficult to promote in clinical practice.

[0003] Deep learning technology excels in high-dimensional data processing and nonlinear fitting, but its application in TCP / NTCP model construction is still in its early stages, with key technical shortcomings: a dedicated network architecture adapted to radiotherapy dose data has not been developed; the problems of limited clinical data sample size and uneven label distribution remain unresolved; a unified modeling framework integrating heterogeneous data from multiple sources such as dose, clinical data, imaging, and genetic data is lacking; and the "black box" nature of the models results in insufficient interpretability. Therefore, there is an urgent need to construct a deep learning TCP / NTCP dose model and its engineering system that integrates multi-source data, accurately captures dose-effect relationships, and possesses both good generalization performance and interpretability to meet the needs of clinical radiotherapy planning optimization. This has become a pressing technical challenge in this field. Summary of the Invention

[0004] In view of the aforementioned shortcomings of existing technologies, the technical problem to be solved by this invention is the lack of integration of multi-source heterogeneous data, inaccurate capture of dose space features, poor generalization ability due to limited sample size and label imbalance, and lack of clinical interpretability in existing TCP / NTCP models. This invention provides a method and system for generating dose schemes based on deep learning for TCP and NTCP. Through multimodal data fusion and a deep learning architecture, it improves the model's prediction accuracy and generalization ability, achieving synergistic optimization of maximizing TCP and minimizing NTCP within the tolerance range of normal tissues, providing technical support for individualized radiotherapy dose planning.

[0005] To achieve the above objectives, this invention provides a method for generating dose schemes using TCP and NTCP based on deep learning, comprising the following steps:

[0006] Acquire multi-source heterogeneous data from the target patient population;

[0007] Multi-source heterogeneous data is processed to form a standardized multimodal feature dataset with consistent dimensions;

[0008] Based on the follow-up results of the target patient group, the corresponding TCP and NTCP labels are labeled for each sample in the standardized multimodal feature dataset;

[0009] A TCP-NTCP joint prediction model based on deep learning is constructed and trained using a labeled dataset to enable the model to learn the nonlinear correlation between multimodal data and TCP and NTCP.

[0010] The trained model is applied to the target patient. By adjusting the input dose parameters, the TCP dose-response curve and NTCP dose-response curve for the individual patient are generated. The TCP-NTCP joint dose-response curve is established based on the TCP-NTCP joint dose-response function.

[0011] Based on individual dose-response curves, we establish and derive the optimal screening constraint formula for therapeutic dose regimens. Under the condition of satisfying NTCP clinical constraints, we solve for the optimal dose distribution scheme that maximizes TCP.

[0012] Furthermore, the source heterogeneous data includes tumor dose distribution data, medical imaging data, clinical baseline data, and biomarker data; among which, tumor dose distribution data includes complete raw data of three-dimensional spatial dose distribution of tumor and surrounding normal tissue, prescription dose of target area, and radiation dose data of each organ at risk;

[0013] Medical imaging data includes CT imaging data;

[0014] Clinical baseline data includes basic information such as age and whether there are underlying diseases;

[0015] Biological marker data specifically include tumor markers and normal tissue tolerance markers;

[0016] Among them, the dose-volume histogram (DVH) feature vector is obtained based on tumor dose distribution data.

[0017] Furthermore, the multi-source heterogeneous data is processed, specifically including...

[0018] The patient's complete three-dimensional spatial dose distribution raw data of tumor and surrounding normal tissue is spatially aligned with medical images to obtain three-dimensional dose matrix data;

[0019] The patient's three-dimensional dose matrix data and dose-volume histogram (DVH) feature vector, as well as the patient's pre-radiotherapy clinical baseline data and biological marker data, are processed to obtain a standardized multimodal feature dataset.

[0020] Furthermore, a deep learning-based TCP-NTCP joint prediction model is constructed, including the following steps:

[0021] Multimodal features are encoded using a dual-branch structure. 3D-CNN extracts spatial features of three-dimensional dose / image, and stacked fully connected networks extract numerical features of DVH and clinical baseline, outputting a unified feature vector of two dimensions.

[0022] By fusing cross-modal features, the dimensions of spatial feature vectors and numerical feature vectors are first aligned, and then the correlation strength between the two types of feature vectors and the prediction target is calculated through learnable weights. The weighted sum is then used to generate a global fused feature vector.

[0023] Based on the dual branches of "feature sharing + task independence", it shares the global fusion feature vector and outputs the normalized prediction probabilities of TCP and NTCP in parallel.

[0024] Define a joint loss function, based on cross-entropy loss, and construct a dual-task loss function by combining L2 weight decay term to balance the optimization priorities of the two prediction tasks;

[0025] To optimize model training, the dataset is divided into a 7:2:1 ratio. The Adam optimizer is used to iteratively update the parameters, and cosine annealing is used to adjust the learning rate. The validation set loss is used as an early stopping condition, and the optimal model parameters are saved.

[0026] Furthermore, the TCP-NTCP joint prediction model includes a multimodal feature encoding unit, a cross-modal fusion unit, and a dual-task prediction unit. The multimodal feature encoding unit extracts key information from each modality of data, which is then integrated by the cross-modal fusion unit. Finally, the dual-task prediction unit learns the nonlinear correlation between the multimodal data and TCP and NTCP, respectively.

[0027] Furthermore, the TCP-NTCP joint dose-response function is... ;in, Indicates the core optimization metrics, The radiation therapy dosage plan includes the prescribed dose to the target area, dose constraints for organs at risk, and dose distribution patterns. : indicates in the dosage regimen The probability of tumor control is as follows. : indicates in the dosage regimen The probability of complications in normal tissues.

[0028] Furthermore, the formula for the optimal screening constraint of the radiotherapy dosage regimen is as follows: ;

[0029] in, ;

[0030]

[0031] This represents the i-th candidate radiotherapy dose regimen, which includes the target volume prescription dose, dose constraints for organs at risk, and dose distribution pattern. Represents the i-th dosage regimen Synergistic benefit metrics; Indicates the dosage regimen from Adjust to At that time, the increase in the probability of tumor control; Indicates the dosage regimen from Adjust to At that time, the increase in the probability of complications in normal tissues; Indicators of synergistic benefits The maximum dosage regimen.

[0032] In a preferred embodiment of the present invention, a system for generating dose using TCP and NTCP based on deep learning is provided, comprising the following modules:

[0033] The data acquisition and labeling module is used to acquire multi-source heterogeneous data of the target patient group and label the data with TCP and NTCP tags based on the follow-up results of the target patient group.

[0034] The data standardization processing module is used to process multi-source heterogeneous data in a unified manner to form a standardized multimodal feature dataset with consistent dimensions.

[0035] The TCP-NTCP joint prediction model module has a built-in deep learning-based TCP-NTCP joint prediction model. It receives labeled, standardized multimodal feature datasets and learns the nonlinear correlation between multimodal data and TCP and NTCP through training.

[0036] The dose-response curve generation module is used to apply the trained model to the target patient. By adjusting the input dose parameters, it generates the TCP dose-response curve and NTCP dose-response curve for the individual patient, and establishes the TCP-NTCP joint dose-response curve based on the TCP-NTCP joint dose-response function.

[0037] The dose optimization solution module is used to establish and derive the optimal screening constraint formula for the therapeutic dose regimen based on individualized dose-response curves, and solve for the optimal dose distribution scheme that maximizes TCP under the condition of satisfying NTCP clinical constraints.

[0038] In another preferred embodiment of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above method.

[0039] In another preferred embodiment of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, performs the steps of the method described above.

[0040] The present invention provides a method and system for generating dose schemes based on deep learning for TCP and NTCP, which has the following beneficial effects:

[0041] First, this invention breaks down the data compatibility barriers of traditional models by standardizing the processing of diverse data, including tumor dose distribution, medical imaging, clinical baselines, and biological markers, enabling the fusion and utilization of multi-dimensional data. Compared to the limitations of traditional models with their single data source, this invention expands the data dimensions, comprehensively capturing key factors affecting TCP / NTCP and laying the foundation for accurate prediction.

[0042] Second, this invention employs deep learning to construct a joint prediction model. Leveraging its strong nonlinear fitting capabilities, it accurately uncovers the intrinsic correlation between multi-source data and TCP / NTCP. Compared to traditional models that struggle to characterize nonlinear relationships and have limited prediction accuracy, this invention can output more accurate prediction results under multiple doses. Training with multi-source data enhances generalization ability, adapting to different patient groups and reducing individual variability bias.

[0043] Third, this invention substitutes the predicted results into the combined dose-response function to construct a curve, and solves for the optimal dose by combining constraints, achieving the goal of "maximizing TCP and minimizing NTCP within the tolerance of normal tissue". Compared with traditional models that consider TCP or NTCP alone, which can easily lead to treatment imbalance, this invention takes into account both tumor control and normal tissue protection, provides a quantitative basis for optimization, formulates personalized dosage plans, and improves the targeting of the plan.

[0044] Fourth, the technical solution of this invention aligns with clinical needs, and the entire process from data acquisition and follow-up annotation to dose calculation is consistent with the logic of diagnosis and treatment. Compared with the shortcomings of traditional model predictions that are disconnected from clinical practice, the optimal dose value can directly provide a precise reference for individualized planning, helping physicians balance efficacy and safety, reduce complications, improve control rates, and promote the transformation of radiotherapy from experience-based to precision-based, demonstrating significant clinical value.

[0045] Fifth, this invention achieves efficient connection between data input and dose output through an integrated process of data standardization, deep learning prediction, curve construction, and dose calculation. Compared to the complex process of traditional manual data integration and repeated trial calculations, it simplifies operation steps, reduces reliance on physician experience, quickly provides optimized solutions, improves planning efficiency, and facilitates the widespread clinical application of precision radiotherapy.

[0046] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a preferred embodiment of the method for generating dose schemes using TCP and NTCP based on deep learning.

[0048] Figure 2 This is a network architecture diagram of a TCP-NTCP joint prediction model for a method of generating dose schemes based on deep learning, according to a preferred embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram of the system framework of a method for generating dose using TCP and NTCP based on deep learning, which is a preferred embodiment of the present invention.

[0050] Figure 4 This is a diagram showing the individualized TCP-NTCP dose-response curve and joint optimization results of a preferred embodiment of the present invention;

[0051] Figure 5 This is a visualization of the optimal screening constraints for radiotherapy dosage schemes according to a preferred embodiment of the present invention. Detailed Implementation

[0052] 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.

[0053] 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.

[0054] like Figure 1 As shown, this embodiment of the invention provides a method for generating dose schemes using TCP and NTCP based on deep learning, including steps 101-105, as follows:

[0055] Step 101: Obtain multi-source heterogeneous data of the target patient population; obtain patient tumor dose distribution, medical imaging, clinical baseline and biomarker data.

[0056] Step 102: Standardize the multivariate data to obtain a dataset with consistent dimensions, and label it with corresponding TCP and NTCP tags based on patient follow-up results. Specifically, process the multi-source heterogeneous data to form a standardized multimodal feature dataset with consistent dimensions; based on the follow-up results of the target patient group, label each sample in the standardized multimodal feature dataset with corresponding TCP and NTCP tags.

[0057] Step 103: Construct a TCP-NTCP joint prediction model based on deep learning, train it using the labeled dataset, and enable the model to learn the nonlinear relationship between multimodal data and TCP and NTCP; input the training dataset into the TCP-NTCP joint prediction model based on deep learning, learn the nonlinear relationship between data and TCP / NTCP, and output the prediction results under different doses;

[0058] Step 104: Apply the trained model to the target patient. By adjusting the input dose parameters, generate the TCP dose-response curve and NTCP dose-response curve for the individual patient. Establish the TCP-NTCP joint dose-response curve based on the TCP-NTCP joint dose-response function.

[0059] Substituting the prediction results into the TCP-NTCP joint dose-response function, a joint dose-response curve is established. The formula for the TCP-NTCP joint dose-response function is:

[0060] ;

[0061] in, Dosage regimen The corresponding collaborative optimization indicators, This represents the predicted probability of tumor control at dose G. For dosage Predicted probability values ​​of complications in normal tissues;

[0062] Step 105: Based on the individual dose-response curve, establish and derive the optimal screening constraint formula for the radiotherapy dosage regimen. Under the condition of satisfying the NTCP clinical constraints, solve for the optimal dose distribution scheme that maximizes TCP. Then, substitute the optimal screening constraint formula for the radiotherapy dosage regimen into the optimal screening constraint formula to solve for the optimal dose value that balances tumor control and normal tissue protection. The optimal screening constraint formula for the radiotherapy dosage regimen is:

[0063] ;

[0064] in, ;

[0065] in,

[0066] This represents the i-th candidate radiotherapy dose regimen, which includes the target volume prescription dose, dose constraints for organs at risk, and dose distribution pattern. Represents the i-th dosage regimen Synergistic benefit metrics; Indicates the dosage regimen from Adjust to At that time, the increase in the probability of tumor control; Indicates the dosage regimen from Adjust to At that time, the increase in the probability of complications in normal tissues; Indicators of synergistic benefits The highest dose regimen. Select from the effective regimens. The largest dose distribution scheme is used as the optimal dose distribution scheme to ensure that NTCP clinical constraints are met while TCP is maximized.

[0067] Step 103, which involves constructing a deep learning-based TCP-NTCP joint prediction model, specifically includes the following steps:

[0068] The multimodal feature coding unit consists of two parallel modules: a spatial feature coding subunit and a numerical feature coding subunit, which process different types of input data respectively.

[0069] The spatial feature encoding subunit adopts a 3D-CNN network architecture, and the input is tensor data fused from a three-dimensional dose matrix and medical images. ,in These represent the height, width, and depth of the tensor, respectively. The core operation of the network is 3D convolution, and its calculation formula is:

[0070]

[0071] In the formula, Input features to the convolutional layer, 3D convolution kernel, For bias terms, The spatial coordinates of the characteristic tensor This is the size index of the convolutional kernel. The output of the convolutional layer undergoes a nonlinear transformation via the ReLU activation function, as shown in the formula: Then, downsampling is performed using the MaxPool3d pooling layer, calculated as follows:

[0072]

[0073] In the formula, This represents the pooling stride. After multiple levels of convolution and pooling, the feature tensor is compressed into a one-dimensional vector using a global average pooling layer, resulting in a final output dimension of [dimensionality missing]. Spatial feature vectors .

[0074] The numerical feature encoding subunit adopts a three-level stacked "Linear→ReLU→Dropout" fully connected network structure. The input is the DVH feature vector, the numerical tensor obtained by concatenating clinical baseline data and biological marker data. ,in This represents the total number of numerical features. The calculation formula for each fully connected layer is:

[0075]

[0076] In the formula, This is the weight matrix. This is the bias term. After ReLU activation, the output is processed by a Dropout layer that randomly discards some neurons to suppress overfitting; the probability of Dropout is set to 0.2. After processing by a three-level fully connected network, the output dimension is... Numerical eigenvectors .

[0077] Constructing cross-modal fusion units

[0078] First, consider the spatial feature vectors. With numerical eigenvectors Dimension alignment is performed using a fully connected layer. Mapping to For the same dimension, the mapping formula is:

[0079]

[0080] In the formula, To align the weight matrix, For the alignment bias term, the mapped result is .

[0081] Then a learnable weight matrix is ​​introduced. Calculate the association strength between the two types of features and the prediction target, and generate an adaptive weight vector. The calculation formula is:

[0082]

[0083]

[0084] In the formula, Weight matrices Two row vectors.

[0085] Finally, the two types of feature vectors are weighted and summed based on the weight vector to obtain the global fused feature vector, as shown in the formula:

[0086]

[0087] In the formula, This is the final output global fusion feature vector.

[0088] Build a dual-task prediction unit

[0089] It adopts a two-branch architecture of "feature sharing + task independence", with the two branches sharing the global fusion feature vector. As input, while keeping network parameters independent. Each branch contains two levels of fully connected layers, the first level of which will... Mapping to hidden layer features The second fully connected layer maps the hidden layer features to one-dimensional predicted values.

[0090] The predicted values ​​are normalized to the Sigmoid activation function. The normalized prediction probabilities of TCP and NTCP are obtained from the interval, and the calculation formulas are as follows:

[0091]

[0092]

[0093] In the formula, Here are the weight matrices for each fully connected layer branch. For the corresponding bias term, The Sigmoid function has the following formula: , These are the predicted probabilities for TCP and NTCP, respectively.

[0094] Define the joint loss function for the two tasks.

[0095] Based on cross-entropy loss, a joint loss function for two tasks is constructed to achieve collaborative optimization of TCP and NTCP prediction tasks. The formula for calculating the cross-entropy loss of a single task is as follows:

[0096]

[0097] In the formula, This refers to the batch sample size. For the sample The true label, For the sample The predicted probability.

[0098] Based on the above formula, the loss function for the TCP prediction task is: The loss function for the NTCP prediction task is: The formula for calculating the joint loss function for the two tasks is:

[0099]

[0100] In the formula, This is the L2 weight decay coefficient, with a value of [value missing]. L2 corresponds to the L2 norm (also known as the Euclidean norm), which is a mathematical metric for measuring the size of a parameter vector. This is the sum of the L2 norms of all network weight parameters, used to further suppress model overfitting.

[0101] Model training and optimization

[0102] The first step is dataset partitioning: the labeled, standardized multimodal feature dataset is divided into... The model is divided into training, validation, and test sets. The training set is used for iterative updates of model parameters, the validation set is used to monitor the model's generalization ability, and the test set is used to evaluate the model's final performance.

[0103] The second step is optimizer configuration: The Adam optimizer is selected to iteratively update the model parameters. The parameter update formula is as follows:

[0104]

[0105]

[0106]

[0107]

[0108] In the formula, These are the first-order moment estimate and the second-order moment estimate of the gradient, respectively. For the corrected moment estimate, To estimate the attenuation coefficient for moments, The initial learning rate is set to a value of [value to be filled in]. , For numerically stable terms, For the first Weight parameters after round of iteration.

[0109] The third step is learning rate adjustment: The learning rate is dynamically adjusted using a cosine annealing strategy. The calculation formula is as follows:

[0110]

[0111] In the formula, These represent the maximum and minimum values ​​of the learning rate, respectively. This represents the current iteration round number. The maximum number of iterations is [value to be filled in]. .

[0112] Step 4, Training termination condition: Continuous loss on the validation set. Using the absence of a decrease in epochs as an early stopping condition, model training is stopped, and the model parameters with the minimum loss on the validation set are saved, thus completing the construction of the TCP-NTCP joint prediction model.

[0113] like Figure 2 As shown, the TCP-NTCP joint prediction model network architecture diagram of the method for generating dose schemes based on deep learning of TCP and NTCP of the present invention includes a multimodal feature encoding unit, a cross-modal fusion unit, and a dual-task prediction unit. The multimodal feature encoding unit extracts key information of each modality data, which is then integrated by the cross-modal fusion unit. The dual-task prediction unit learns the nonlinear correlation between multimodal data and TCP and NTCP respectively.

[0114] The core function of the multimodal feature coding unit is to adapt to the feature characteristics of heterogeneous data and extract features for different types of data. Specifically, it is divided into spatial feature coding subunits and numerical feature coding subunits.

[0115] Specifically, the spatial feature encoding subunit employs a 3D-CNN network for feature extraction. The 3D-CNN network captures the local correlations and overall distribution features of data in three-dimensional space through convolutional operations, introduces nonlinear transformations using the ReLU activation function, performs feature downsampling and dimensionality compression through the MaxPool3d layer, and finally integrates global spatial information through a global average pooling layer. This accurately extracts the spatial correlation features between "dose distribution and anatomical structure," ultimately outputting a uniform 1×256 spatial feature vector, providing a suitable spatial feature foundation for subsequent cross-modal fusion. The numerical feature encoding subunit: For numerical data such as DVH vectors, clinical baselines, and biological markers, a fully connected network (with a three-level stacked structure of "Linear→ReLU→Dropout") is used for encoding. In the treatment embodiment for target tumor patients, this subunit can extract high-dimensional abstract features from the data, outputting a 1×32 numerical feature vector.

[0116] The numerical feature encoding subunit targets numerical data without spatial attributes, such as DVH vectors, clinical baselines, and biological markers, and employs a fully connected network to complete feature encoding. The fully connected network can uncover linear and nonlinear relationships between numerical data through hierarchical mapping, enhance feature representation through the ReLU activation function, and introduce a Dropout layer to suppress model overfitting. Through a three-level stacked network structure, it progressively extracts high-dimensional abstract features from the data, ultimately outputting a 1×32-dimensional numerical feature vector, achieving effective encoding and dimensionality unification of numerical data.

[0117] The cross-modal fusion unit first performs dimension alignment on the feature vectors output from the sub-modalities. Then, using a learnable weight matrix, it calculates the correlation between the aligned spatial and numerical feature vectors, automatically learning the association strength between different modal features and TCP / NTCP prediction targets, thereby generating corresponding feature weights. Spatial features, directly related to the core information of dose distribution and anatomical structure, have a weight of approximately 0.85, while numerical features have a weight of approximately 0.15, automatically highlighting key prediction information. Finally, based on the generated feature weights, a weighted summation is performed on the two types of aligned feature vectors to obtain a global fusion feature vector with a dimension of 1×256.

[0118] The dual-task prediction unit is based on a two-branch architecture of "feature sharing + task independence," ensuring the independence of the two types of probability predictions and making full use of the complementary information of the globally fused features to simultaneously predict tumor control and normal tissue risk. Specifically, this unit adopts a fully connected network structure with two independent branches. The two branches share the input of the globally fused feature vector, but maintain independent training of network parameters—avoiding interference between TCP and NTCP prediction tasks, while allowing both types of tasks to fully reuse the effective information after multimodal fusion.

[0119] The TCP prediction branch and the NTCP prediction branch have the same network logic. Both use multi-level fully connected layers to gradually map the 1×256 global fusion feature vector to the target prediction space. The ReLU activation function is used to introduce nonlinear transformation to enhance the feature expression ability. Finally, the Sigmoid activation function is used to normalize the output results to the [0,1] interval, so as to obtain the TCP prediction probability and NTCP prediction probability under the corresponding dose.

[0120] During the model training phase, to achieve synergistic optimization of the two prediction tasks, this unit employs a dual-task joint loss function for training. This loss function balances the training priorities of the two tasks by assigning equal weights to the TCP prediction loss and the NTCP prediction loss and summing them, avoiding model bias towards single-task optimization and ensuring that the prediction accuracy of TCP and NTCP improves synchronously. The TCP prediction loss is calculated from the prediction probability and the true tumor control label, while the NTCP prediction loss is calculated from the prediction probability and the true complication label.

[0121] like Figure 3 As shown, this embodiment of the invention provides a system for generating dose schemes based on deep learning for TCP and NTCP, comprising:

[0122] The data acquisition and labeling module is used to acquire multi-source heterogeneous data of the target patient group and label the data with TCP and NTCP tags based on the follow-up results of the target patient group; the multi-source heterogeneous data includes at least the patient's tumor dose distribution data, medical imaging data, clinical baseline data and biomarker data;

[0123] The data standardization module is used to uniformly process multi-source heterogeneous data to form a standardized multimodal feature dataset with consistent dimensions. For spatial data such as three-dimensional dose matrices and medical images, rigid registration is first used to achieve spatial alignment between anatomical structures and dose distributions, and then interpolation is used to unify the data dimensions. For numerical data such as DVH vectors, clinical baselines, and biological markers, min-max normalization is used to map features to the [0,1] interval to eliminate scale differences.

[0124] The TCP-NTCP joint prediction model module incorporates a deep learning-based TCP-NTCP joint prediction model. It receives labeled, standardized multimodal feature datasets and trains to learn the non-linear correlation between multimodal data and TCP / NTCP. This module is the core algorithm of the system. It receives the labeled dataset output from the data standardization module and trains deep learning to learn the non-linear correlation between multimodal features and TCP / NTCP, achieving the mapping capability of "input multimodal features → output corresponding dose TCP / NTCP prediction values".

[0125] The dose-response curve generation module is used to apply the trained model to the target patient. By adjusting the input dose parameters, it generates the TCP dose-response curve and NTCP dose-response curve for the individual patient, and establishes the TCP-NTCP joint dose-response curve based on the TCP-NTCP joint dose-response function.

[0126] The dose optimization solution module is used to establish and derive the optimal screening constraint formula for the therapeutic dose regimen based on individualized dose-response curves, and solve for the optimal dose distribution scheme that maximizes TCP under the condition of satisfying NTCP clinical constraints.

[0127] Among them, the data acquisition and labeling module includes the raw data of three-dimensional spatial dose distribution of the patient's tumor and surrounding normal tissues, the prescription dose of the target area, and the irradiation dose data of each organ at risk. Based on this data, a dose-volume histogram (DVH) feature vector is generated, which includes 20 core features such as average dose, maximum dose, and volume percentage dose.

[0128] Medical imaging data includes acquiring chest CT images of patients before radiotherapy, delineating the tumor target area and the contours of organs at risk using medical imaging processing software, extracting radiomics features, and converting them into tensor format.

[0129] Clinical baseline data includes information collected on the patient's age, gender, whether they have underlying diseases such as hypertension, diabetes, and chronic lung disease, tumor pathology type, and clinical stage.

[0130] The biomarker data includes the detection results of peripheral blood tumor markers and normal tissue tolerance markers collected from patients before radiotherapy. After standardizing the detection scale, core indicators are extracted and standardized feature vectors are generated.

[0131] In addition to the collection of multi-source heterogeneous data, the follow-up data collection nodes were set at 1 month, 3 months, 6 months, 12 months and 24 months after radiotherapy. Data were obtained through imaging examinations, laboratory tests and clinical sign assessments. Label determination must be based on clinical standards: TCP label is based on "no tumor recurrence / metastasis within 24 months" (marked "1" to represent tumor control and "0" to represent no control), and NTCP label is based on the Common Terminology Standard for Adverse Events (CTCAE) (complications of grade III and above are marked "1", and those below are marked "0"), to ensure that the labels are consistent with the actual clinical efficacy and risks.

[0132] like Figure 4 As shown, the individualized TCP-NTCP dose-response curve and joint optimization results provided in this embodiment of the invention are illustrated. The TCP-NTCP dose-response function is:

[0133] .

[0134] : Represents the probability of tumor control (breast cancer in this example), which increases monotonically with increasing radiotherapy dose (horizontal axis, unit Gy), reflecting the enhanced killing effect of increased dose on tumor cells.

[0135] This represents the probability of complications in normal tissue (breast tissue in this case), and it increases slowly with increasing dose.

[0136] : is the difference between "tumor control benefits" and "normal tissue damage risk", i.e., the synergistic optimization index.

[0137] like Figure 4 As shown, The clinical tolerance threshold representing NTCP is the upper limit of the acceptable risk of complications in normal tissues. When the concentration reaches 0.3, the risk of complications in normal tissues is at a clinically tolerable borderline level. A dose exceeding 0.3 (e.g., 0.4, 0.5) indicates an excessively high risk of complications, and clinicians will carefully select or not recommend using the corresponding radiotherapy dose.

[0138] First, the single-fraction dose for tumor radiotherapy was determined to be 2 Gy. The total radiotherapy dose was then discretely divided into 2 Gy units to generate a fractionated irradiation dose plan. (20, 22, 24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52, 54, 56, 58, 60, 62, 64, 66, 68, 70 Gy, step size 2 Gy).

[0139] Then, based on the pre-defined TCP-NTCP joint prediction model, the dosage regimens are calculated separately. Corresponding tumor control probability Probability of complications compared to normal tissue And simultaneously calculate collaborative optimization indicators. That is, the difference between the benefits of tumor control and the risk of damage to normal tissues.

[0140] Therefore, when =0.05, =0.05, =0.05-0.05=0.00, the dosage is too low, the tumor control effect is extremely poor, although there is no risk of complications, the net benefit is extremely low, and there is no clinical treatment value; when =0.83,, =0.12, then =0.83-0.12=0.71, the dose enters the effective therapeutic window, the tumor control rate has reached a high level, and the risk of normal tissue loss is negligible, resulting in a significant increase in net benefit; when =0.95, =0.39, then =0.95-0.39=0.56, TCP has no room for improvement, but NTCP has reached the clinically tolerable threshold, and the net benefit is slightly lower than the optimal dose, so it should be used with caution; when =0.95 (no growth) =0.47, then =0.95-0.47=0.48, NTCP exceeds the clinical tolerance threshold of 0.3, the risk of complications is significantly increased, but the tumor control effect is not improved, the net benefit is significantly reduced, and clinical use is not recommended.

[0141] It should be noted that, in this embodiment of the invention, the TCP-NTCP dose-response curve is obtained through the TCP-NTCP dose-response function, and the combined optimization result is obtained by combining it with the clinical tolerance threshold of NTCP.

[0142] like Figure 5 As shown, this is a visualization of the optimal screening constraints for radiotherapy dosage schemes provided in this embodiment of the invention; wherein, the optimal screening constraint formula for radiotherapy dosage schemes is:

[0143] .

[0144] First, the single-fraction dose for tumor radiotherapy was determined to be 2 Gy. The total radiotherapy dose was then discretely divided into 2 Gy units to generate a fractionated irradiation dose plan. (20, 22, 24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52, 54, 56, 58, 60, 62, 64, 66, 68, 70 Gy, step size 2 Gy).

[0145] Determine the dosage regimen for fractionated irradiation. Subsequently, based on the TCP-NTCP dose-response curves and joint optimization results, the following was obtained: .

[0146] Subsequently, the TCP-NTCP joint prediction model was used to calculate the predictions for every two adjacent... Corresponding "tumor control probability increment" " ;) and "increase in the probability of complications in normal tissues" " ), and simultaneously calculate each .

[0147] Constraint meaning: This means that "the tumor control gain from increasing the unit dose is less than the increase in the risk of normal tissue damage"—at this point, further increasing the dose would lead to an imbalance between risk and benefit, therefore it is necessary to screen for drugs that meet this constraint. .

[0148] Subsequently, clinical tolerance thresholds based on NTCP were established. Filter out those that satisfy the constraint .

[0149] Valid candidate solutions Distribution: satisfying the above two constraints For discrete dose schemes with step sizes of 2 Gy within the range of 48, 50, 52, 54, 56, 58 Gy, and within the range of 48-58 Gy, the corresponding dose schemes are... The trend decreases sequentially, among which =50Gy corresponds to =0.809 is the maximum value; optimal dose selection: among the effective candidate regimens, =50Gy corresponds to =0.809 is the maximum value, therefore this dose is labeled as the "optimal dose". =50Gy". This dosage meets the requirements. The constraints, without reaching the clinical tolerance threshold of NTCP, have been achieved. Maximize.

[0150] For example, when , , The constraints are satisfied; Gy meets the criteria;

[0151] , , , , < Satisfying constraints Gy meets the criteria.

[0152] This method possesses forward-looking risk warning and pre-planning optimization capabilities. Through precise quantitative prediction of TCP / NTCP using a deep learning model, it can identify potential dose risk thresholds (such as the tolerance dose threshold of a critical organ or the ineffective dose range for tumor control) in the early stages of radiotherapy planning. This provides physicians with quantitative evidence for "early risk avoidance + precise dose allocation." This feature avoids the iterative modification process of "design first, then evaluate" in traditional planning, reducing the potential damage of ineffective doses to normal tissues while ensuring effective coverage of the tumor target area. This allows radiotherapy planning to have the dual guarantee of "precise tumor control + safe tissue protection" from the source, further enhancing the scientific rigor and forward-looking nature of treatment plans.

[0153] This system boasts low barriers to clinical implementation and user-friendliness. Its design aligns with the routine workflow of radiotherapy departments, featuring a simple and intuitive visual interface and standardized data interfaces, allowing physicians to quickly learn and use it without needing complex deep learning techniques or data processing knowledge. Furthermore, the system is seamlessly compatible with existing radiotherapy planning systems (TPS), medical imaging systems (PACS), and other commonly used clinical software, eliminating the need to reconstruct the existing digital healthcare ecosystem and significantly reducing the time and cost of data migration and system adaptation. Without disrupting the routine clinical treatment process, the system efficiently outputs optimal dosage plans, accelerating the translation of precision radiotherapy technology from research to clinical practice, and enabling more medical institutions to quickly benefit from the improved treatment outcomes brought about by technological advancements.

[0154] The embodiments of the present invention achieve both the relative maximization of the probability of tumor control and the control of the risk of complications in normal tissues within a clinically acceptable range, thus solving the problem of "difficulty in balancing tumor control and safety" in traditional empirical dosing planning.

[0155] This application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement the steps in the above-described embodiment of the TCP and NTCP dose model construction method based on deep learning, which will not be repeated here.

[0156] This application provides a computer program product that, when run on the hardware system described in this invention (such as a single medical-grade computer or a deep learning server), enables the hardware system to implement the steps described in the various method embodiments above.

[0157] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to the hardware system of this invention, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0158] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for generating dose schemes using TCP and NTCP based on deep learning, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data from the target patient population; The multi-source heterogeneous data is processed to form a standardized multimodal feature dataset with consistent dimensions; Based on the follow-up results of the target patient group, each sample in the standardized multimodal feature dataset is labeled with corresponding TCP and NTCP labels; A TCP-NTCP joint prediction model based on deep learning is constructed and trained using a labeled dataset to enable the model to learn the nonlinear correlation between multimodal data and TCP and NTCP. The trained model is applied to the target patient. By adjusting the input dose parameters, the TCP dose-response curve and NTCP dose-response curve for the individual patient are generated. The TCP-NTCP joint dose-response curve is established based on the TCP-NTCP joint dose-response function. Based on individual dose-response curves, we establish and derive the optimal screening constraint formula for therapeutic dose regimens. Under the condition of satisfying NTCP clinical constraints, we solve for the optimal dose distribution scheme that maximizes TCP.

2. The method for generating dose schemes based on deep learning for TCP and NTCP as described in claim 1, characterized in that, The heterogeneous source data includes tumor dose distribution data, medical imaging data, clinical baseline data, and biomarker data; wherein, the tumor dose distribution data includes complete raw data of three-dimensional spatial dose distribution of the tumor and surrounding normal tissues, prescription dose of the target area, and radiation dose data of each organ at risk; The medical imaging data includes CT imaging data; The clinical baseline data includes basic information such as age and whether there are underlying diseases. The biomarker data specifically includes tumor markers and normal tissue tolerance markers; Among them, the dose-volume histogram (DVH) feature vector is obtained based on tumor dose distribution data.

3. The method for generating dose schemes based on deep learning for TCP and NTCP as described in claim 1, characterized in that, The processing of the multi-source heterogeneous data specifically includes: The patient's complete three-dimensional spatial dose distribution raw data of tumor and surrounding normal tissue is spatially aligned with medical images to obtain three-dimensional dose matrix data; The patient's three-dimensional dose matrix data and dose-volume histogram (DVH) feature vector, as well as the patient's pre-radiotherapy clinical baseline data and biological marker data, are processed to obtain a standardized multimodal feature dataset.

4. The method for generating dose schemes based on deep learning for TCP and NTCP as described in claim 1, characterized in that, Constructing a deep learning-based TCP-NTCP joint prediction model includes the following steps: Multimodal features are encoded using a dual-branch structure. 3D-CNN extracts spatial features of three-dimensional dose / image, and stacked fully connected networks extract numerical features of DVH and clinical baseline, outputting a unified feature vector of two dimensions. By fusing cross-modal features, the dimensions of spatial feature vectors and numerical feature vectors are first aligned, and then the correlation strength between the two types of feature vectors and the prediction target is calculated through learnable weights. The weighted sum is then used to generate a global fused feature vector. Based on the dual branches of "feature sharing + task independence", it shares the global fusion feature vector and outputs the normalized prediction probabilities of TCP and NTCP in parallel. Define a joint loss function, based on cross-entropy loss, and construct a dual-task loss function by combining L2 weight decay term to balance the optimization priorities of the two prediction tasks; To optimize model training, the dataset is divided into a 7:2:1 ratio. The Adam optimizer is used to iteratively update the parameters, and cosine annealing is used to adjust the learning rate. The validation set loss is used as an early stopping condition, and the optimal model parameters are saved.

5. The method for generating a dose scheme based on deep learning for TCP and NTCP as described in claim 4, characterized in that, The TCP-NTCP joint prediction model includes a multimodal feature encoding unit, a cross-modal fusion unit, and a dual-task prediction unit. The multimodal feature encoding unit extracts key information from each modality of data, which is then integrated by the cross-modal fusion unit. Finally, the dual-task prediction unit learns the nonlinear correlation between the multimodal data and TCP and NTCP.

6. The method for generating dose schemes based on deep learning for TCP and NTCP as described in claim 1, characterized in that, The TCP-NTCP joint dose-response function is ;in, Indicates the core optimization metrics, This indicates the radiotherapy dosage regimen, including the prescription dose to the target area, dose constraints for organs at risk, and dose distribution patterns. : indicates in the dosage regimen The probability of tumor control is as follows. : indicates in the dosage regimen The probability of complications in normal tissues.

7. The method for generating a dose scheme based on deep learning for TCP and NTCP as described in claim 1, characterized in that, The formula for the optimal screening constraints of radiotherapy dosage regimens is as follows: ; in, ; ， This represents the i-th candidate radiotherapy dose regimen, which includes the target volume prescription dose, dose constraints for organs at risk, and dose distribution pattern. Represents the i-th dosage regimen Synergistic benefit metrics; Indicates the dosage regimen from Adjust to At that time, the increase in the probability of tumor control; Indicates the dosage regimen from Adjust to At that time, the increase in the probability of complications in normal tissues; Indicators of synergistic benefits The maximum dosage regimen.

8. A system for generating dose using TCP and NTCP based on deep learning, characterized in that, Includes the following modules: The data acquisition and labeling module is used to acquire multi-source heterogeneous data of the target patient group and label the data with TCP and NTCP tags based on the follow-up results of the target patient group. The data standardization processing module is used to perform unified processing on the multi-source heterogeneous data to form a standardized multimodal feature dataset with consistent dimensions. The TCP-NTCP joint prediction model module has a built-in deep learning-based TCP-NTCP joint prediction model. It receives labeled, standardized multimodal feature datasets and learns the nonlinear correlation between multimodal data and TCP and NTCP through training. The dose-response curve generation module is used to apply the trained model to the target patient. By adjusting the input dose parameters, it generates the TCP dose-response curve and NTCP dose-response curve for the individual patient, and establishes the TCP-NTCP joint dose-response curve based on the TCP-NTCP joint dose-response function. The dose optimization solution module is used to establish and derive the optimal screening constraint formula for the therapeutic dose scheme based on the individualized dose-response curve, and solve the optimal dose distribution scheme that maximizes TCP under the condition of satisfying the NTCP clinical constraints.

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 program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

Images