A Method and System for Optimizing Radiotherapy Lattice Parameters Based on Spatial Segmentation Technology

Abstract

A radiotherapy lattice parameter optimization method based on spatial segmentation technology includes the following steps: acquiring the patient's radiotherapy image data; calculating the tumor geometric features based on the radiotherapy image data and preset geometric constraints; inputting the tumor geometric features and initial lattice parameters into a lattice parameter optimization model constructed based on the PPO algorithm to identify and obtain the optimal lattice parameters. The lattice parameter optimization model includes a policy network, a value network, a parameter and optimization module, and a reward calculation module. This invention can output the optimal lattice parameters based on the patient's radiotherapy image data, providing a reliable basis for spatial segmentation radiotherapy, and effectively improving efficiency without relying on human experience.

Description

Technical Field

[0001] This invention relates to the field of lattice parameter optimization technology for spatial segmentation radiotherapy, specifically a method and system for optimizing radiotherapy lattice parameters based on spatial segmentation technology. Background Technology

[0002] Spatially Fractionated Radiotherapy (SFRT) is a novel radiotherapy method that improves treatment efficacy by generating alternating high and low dose distribution patterns within the tumor. SFRT discretely distributes multiple high-dose lattice target regions within the tumor to create a dose "peak-valley" distribution; that is, high-dose peaks are formed in the regions where the lattice is located, while low-dose valleys are formed in the tumor areas not covered by the lattice. Studies have shown that SFRT can increase the peak dose within the tumor without significantly increasing damage to normal tissues, and can activate the immune response within the tumor, achieving a higher local tumor control rate. SFRT has shown good efficacy, especially for tumors that are difficult to treat with conventional radiotherapy (such as large, well-tolerated tumors).

[0003] However, a key challenge in the practical application of SFRT is how to rationally set and optimize lattice parameters, such as lattice spacing, radius, and peak-to-valley dose ratio (PVDR). These parameters directly affect the physician's choice of treatment method, but because each patient's tumor size and morphology are different, manual or static algorithm-based parameter settings are often difficult to adapt to individual needs. If the lattice distribution is too dense, it may lead to a lower peak-to-valley dose ratio, increasing damage to normal tissues; if the lattice distribution is too sparse, it may not be able to effectively cover the entire tumor area. Therefore, to achieve the optimal dose distribution, a lattice optimization method that can adaptively adjust according to the specific tumor morphology of the patient is needed.

[0004] In existing technologies, SFRT primarily relies on operators manually positioning and sizing the lattice. Typically, radiation planners place the lattice one by one using a two-dimensional image or three-dimensional model to achieve the desired dose peak-to-valley distribution. However, this method is highly dependent on the operator's experience and judgment, resulting in low efficiency and significant errors. When the tumor is large or complex in shape, manual lattice placement is not only time-consuming but also struggles to achieve ideal uniformity, making it difficult to meet the optimal peak-to-valley ratio requirements for dose distribution. Furthermore, because the position and size of manually placed lattices are relatively fixed, operators cannot flexibly adjust lattice parameters during radiation planning, thus reducing the therapeutic effect of SFRT.

[0005] To address the shortcomings of manual placement, some studies have proposed lattice placement methods based on geometric algorithms, such as the closest packing method. One study proposed a scheme using the closest packing algorithm, which determines the lattice spacing and initial position based on the tumor volume and shape, and uses the closest packing principle to arrange multiple spherical lattices so that they cover the tumor target area as uniformly as possible. This method achieves spatial dose segmentation through a uniform geometric distribution and is suitable for tumors with regular shapes and moderate sizes. However, because the parameter settings in geometric algorithms are usually fixed, it is difficult to flexibly handle complex tumor morphologies.

[0006] Other studies have proposed lattice arrangement optimization schemes based on multivariate Gaussian distributions, achieving dose control within the tumor region through iterative optimization with fixed parameters. This method sets fixed lattice parameters (such as spacing and diameter) using multimodal imaging to arrange the lattice within the tumor region, suitable for optimization needs within a certain range, but with limited adaptability under different tumor morphologies.

[0007] With the advancement of artificial intelligence technology, deep learning has been increasingly applied to the prediction and optimization of radiotherapy dose distribution. For example, existing literature has attempted to use convolutional neural networks (CNNs) and generative adversarial networks (GANs) to predict radiotherapy dose distribution, achieving significant results in real-world cases. However, most of these studies focus on traditional radiotherapy with uniform dose distribution and are applied to simple two-dimensional or three-dimensional models, lacking exploration of applications for non-uniform dose distribution in SFRT technology. Furthermore, deep learning models typically rely on large amounts of labeled data for training, while clinical data on non-uniform dose distribution is scarce in the SFRT field, making it difficult to obtain sufficient training data. Therefore, current deep learning-based dose prediction models are not yet fully applicable to lattice parameter optimization in SFRT.

[0008] In summary, the existing SFRT lattice arrangement schemes have the following main problems: manual arrangement is inefficient, fixed parameter settings based on geometric algorithms are difficult to meet personalized needs, and radiotherapy optimization schemes based on deep learning and reinforcement learning have not been fully applied to dose distribution optimization of SFRT. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method and system for optimizing radiotherapy lattice parameters based on spatial segmentation technology. This method can output optimal lattice parameters based on the patient's radiotherapy image data, providing a reliable basis for spatial segmentation radiotherapy. Furthermore, it does not rely on human experience and can effectively improve efficiency.

[0010] To achieve the above objectives, the specific solution adopted by the present invention is as follows: a radiotherapy lattice parameter optimization method based on spatial segmentation technology, comprising the following steps:

[0011] Acquire the patient's radiotherapy imaging data;

[0012] The geometric features of the tumor are calculated based on radiotherapy imaging data and pre-set geometric constraints.

[0013] The tumor geometric features and initial lattice parameters are input into a lattice parameter optimization model built based on the PPO (Proximal Policy Optimization) algorithm. The lattice parameter optimization model includes a policy network, a parameter optimization module, a reward calculation module, and a value network.

[0014] A policy network is used to iteratively generate a lattice optimization strategy based on tumor geometric features and initial lattice parameters.

[0015] The parameter optimization module generates optimized lattice parameters based on a lattice optimization strategy.

[0016] The reward calculation module is used to calculate the reward value of the optimized lattice parameters, and when the reward value meets the preset conditions, the optimized lattice parameters are output as the optimal lattice parameters.

[0017] When the reward value does not meet the preset conditions, the value network is used to evaluate the policy network to obtain the state value. The state value is used to guide the policy network to update the lattice optimization policy through iteration.

[0018] Preferably, the tumor geometric features include tumor volume, shape factor, and boundary features, wherein the shape factor is used to describe the tumor as a regular geometry based on radiotherapy imaging data.

[0019] Preferably, after calculating the geometric features of the tumor, the state space and action space of the lattice parameter optimization model are determined based on the geometric features of the tumor.

[0020] Preferably, the state-space representation of the lattice parameter optimization model is as follows:

[0021] S={Volume, Shape Factor, Boundary Features, Current Parameters};

[0022] Where Volume is the tumor volume, Shape Factor is the shape factor, Boundary Features are the boundary features, and Current Parameters are the current lattice parameter state;

[0023] The action space of the lattice parameter optimization model is represented as:

[0024] A=[Spacing, Radius, Edge distance];

[0025] Where Spacing is the distance between lattice points; Radius is the effective radius of the lattice points; and Edge distance is the distance between the lattice and the edge of the tumor target area.

[0026] Preferred methods for calculating rewards include:

[0027] R=ω1·R PVDR +ω2·R ADR +ω3·R Coverage -ω4·R Penalty ;

[0028] Where R is the reward value, R PVDR R is the peak-to-trough ratio of the dose in radiotherapy. ADR ω1 is the absorbed dose ratio, where R is the dose ratio. PVDR The weights, ω2 is R ADR The weights, and ω1 + ω2 = 1, R Coverage For coverage, R Penalty This is a penalty item.

[0029] Preferably, the peak-to-trough ratio R in radiotherapy PVDR The calculation method is as follows:

[0030]

[0031] Among them, D peak For the average high dose in the lattice target region, D valley The average low dose is in the extralattice region, and D peak and D valley All are determined by optimizing lattice parameters;

[0032] Absorbed dose ratio R ADR The calculation method is as follows:

[0033]

[0034] Among them, D ablative D represents the total dose within the ablation zone. total The total dose to the tumor region, and D ablative and D total All are determined by optimizing the lattice parameters.

[0035] Preferably, the coverage R Coverage The calculation method is as follows:

[0036]

[0037] Wherein, Coverage Area is the volume of the tumor covered by the lattice layout and is determined by optimized lattice parameters, and Tumor Area is the total volume of the tumor and is determined by the tumor geometry.

[0038] Preferably, the penalty term R Penalty The calculation method is as follows:

[0039] R Penalty =R Penalty-Density +R Penalty-Overlap ;

[0040] Among them, R Penalty-Density R is the density penalty factor. Penalty-Overlap This is the overlap penalty factor.

[0041] A radiotherapy lattice parameter optimization system based on spatial segmentation technology includes:

[0042] The input module is used to acquire the patient's radiotherapy image data and to preprocess the radiotherapy image data;

[0043] The tumor identification module is used to calculate the geometric features of the tumor based on radiotherapy image data and preset geometric constraints.

[0044] The lattice parameter optimization module is used to identify the optimal lattice parameters by utilizing the lattice parameter optimization model to identify the geometric features of the tumor.

[0045] Preferably, the system further includes a lattice arrangement and DICOM update module, used to generate lattice arrangement data based on optimal lattice parameters and generate a data transfer file.

[0046] The radiotherapy lattice parameter optimization method of this invention can generate optimal lattice parameters with better peak-to-valley dose ratio (PVDR) and absorbed dose ratio (ADR), providing physicians with more targeted and effective parameter guidance for radiotherapy, especially in the treatment of large or complex-shaped tumors, offering significant assistance to the radiotherapy process. In optimizing lattice parameters, this invention can adaptively adjust the lattice layout parameters according to the specific tumor characteristics of the patient, effectively addressing the dose distribution requirements of irregular and complex-shaped tumors, overcoming the limitations of fixed parameter settings in existing methods, and meeting the specific dose distribution needs of different patients during treatment. Based on deep reinforcement learning technology, this invention can automatically generate optimal lattice parameters from the patient's radiotherapy imaging data, thereby reducing reliance on operator experience and reducing the time and labor costs of manual setup, thus improving the overall efficiency of the treatment process. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the parameter optimization method of the present invention;

[0049] Figure 2 This is a schematic diagram of the structure of the lattice parameter optimization model;

[0050] Figure 3 This is a schematic diagram illustrating the process of lattice parameter change;

[0051] Figure 4 This is a structural block diagram of the lattice parameter optimization system of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1.

[0054] like Figures 1 to 4 As shown, the radiotherapy lattice parameter optimization method based on spatial segmentation technology includes S1 to S7.

[0055] S1. Obtain the patient's radiotherapy imaging data, which mainly includes commonly used medical images such as CT and MRI images. Since radiotherapy is performed after a tumor is diagnosed, and doctors determine the tumor's location in medical images during diagnosis, the acquired radiotherapy imaging data can directly contain the ROI (Region of Interest) information marked by the doctor during the diagnosis process. This facilitates subsequent determination of the tumor's location, size, and other characteristics based on the radiotherapy imaging data.

[0056] S2. Calculate the tumor geometric features based on radiotherapy image data and preset geometric constraints. The geometric constraints are used to simplify the irregular tumor into a regular geometric shape to facilitate subsequent lattice partitioning. In one embodiment of the invention, the tumor geometric features include tumor volume, shape factor, and boundary features, wherein the shape factor is used to describe the tumor as a regular geometric shape based on the radiotherapy image data. More specifically, the shape factor describes the irregular tumor as a regular sphere, thereby calculating the tumor volume. Where r is the radius of the tumor, which can be calculated using graphical fitting based on ROI information. The boundary condition is set to the boundary length, which is the perimeter of a cross section passing through the center of the sphere.

[0057] On the other hand, after calculating the geometric features of the tumor, the state space and action space of the lattice parameter optimization model are determined based on the tumor geometric features. The state space of the lattice parameter optimization model is represented as follows:

[0058] S={Volume, Shape Factor, Boundary Features, Current Parameters};

[0059] Wherein, Volume is the tumor volume, Shape Factor is the shape factor, Boundary Features are the boundary features, and Current Parameters are the current lattice parameter state.

[0060] The action space of the lattice parameter optimization model is represented as:

[0061] A=[Spacing, Radius, Edge distance];

[0062] Spacing is the distance between lattice points; Radius is the effective radius of the lattice points, used to control the coverage of the radiotherapy area; Edge distance is the distance between the lattice and the edge of the tumor target area.

[0063] The state space and action space work together to constrain the lattice parameter optimization model, ensuring that the model can output optimal lattice parameters that better match the tumor geometry and the patient's tumor state under given state and action space conditions. This provides doctors with more accurate and sufficient data references for radiotherapy.

[0064] S3. Input the tumor geometric features and initial lattice parameters into the lattice parameter optimization model constructed based on the PPO (Proximal Policy Optimization) algorithm for identification. The lattice parameter optimization model includes a policy network, a parameter optimization module, a reward calculation module, and a value network.

[0065] S4. A policy network is used to iteratively generate a lattice optimization strategy based on tumor geometry and initial lattice parameters. For example... Figure 1 and 2 As shown, the policy network generates lattice optimization policies and corresponding optimized lattice parameters for optimizing lattice parameters. The policy network uses a multilayer perceptron structure, consisting of four main parts: an input layer, two hidden layers, and an output layer. The input layer receives tumor geometry features and initial lattice parameters input manually. The tumor geometry features include volume, shape, and boundaries, providing the policy network with complete background information for subsequent decision-making. The hidden layer structure consists of two fully connected layers with 256 and 128 nodes respectively, employing the ReLU activation function to enhance the network's nonlinear modeling capabilities. This design captures complex relationships in the lattice layout and efficiently handles optimization problems. The output layer outputs the lattice optimization policy, specifically including lattice spacing, lattice radius, and lattice layout. These parameters directly guide the parameter optimization module in generating optimized lattice parameters. To significantly improve the performance of the policy network, each layer undergoes a nonlinear transformation using the ReLU activation function to enhance its feature extraction capabilities. In the output layer, the tanh function is used to map the results to a reasonable range, specifically the formula a = tanh(W·ReLU(H(s)) + b), where s is the state vector containing tumor geometric features and the current parameter state (i.e., the lattice optimization strategy), a is the output parameter adjustment suggestion, and W and b are the network weights and biases, respectively. In practical applications, the policy network employs a multi-scale optimization strategy, first determining the lattice distribution at a coarse-grained level, and then performing local fine-grained optimization. An experience replay mechanism is also introduced to store historical optimization data for batch training, improving learning effectiveness. Furthermore, randomness can be introduced during parameter search to avoid getting trapped in local optima. The policy network and the value network cooperate through backpropagation, continuously adjusting the lattice optimization strategy generated by the policy network. To avoid long-term ineffective iterations in the policy network and value network, i.e., the lattice optimization strategy does not improve significantly during multiple iterations, this invention also adopts an early stopping mechanism. When the lattice optimization strategy does not show significant differences during multiple consecutive iterations, the iteration is stopped, thereby avoiding wasting computational resources while ensuring the optimization effect.

[0066] S5. Utilize the parameter optimization module to generate optimized lattice parameters based on the lattice optimization strategy. For example... Figure 1 As shown, the parameter optimization module generates optimized lattice parameters based on the lattice optimization strategy. Specifically, the method for generating optimized lattice parameters includes three parts: lattice spacing calculation, lattice radius optimization, and lattice layout adjustment, corresponding to the lattice optimization strategy. More specifically, the strategy network generates a lattice optimization strategy with each iteration, and correspondingly, the parameter optimization module generates an optimized lattice parameter with each iteration. The first generation of optimized lattice parameters is based on manually input initial lattice parameters, which are then optimized using the lattice optimization strategy to generate the first optimized lattice parameter. In subsequent processes, the previously generated optimized lattice parameters are used as the basis, and new lattice optimization strategies are employed to continuously iterate and update the optimized lattice parameters.

[0067] S6. The reward calculation module is used to calculate the reward value for the optimized lattice parameters. When the reward value meets the preset conditions, the optimized lattice parameters are output as the optimal lattice parameters. The reward calculation module is used to evaluate and provide feedback on the optimized lattice parameters in real time through physical model calculations and index evaluation. Specifically, the reward calculation method includes: R = ω1·R PVDR +ω2·R ADR +ω3·R Coverage -ω4·R Penalty ;

[0068] Where R is the reward value, R PVDR R is the peak-to-trough ratio of the dose in radiotherapy. ADR ω1 is the absorbed dose ratio, where R is the dose ratio. PVDR The weights, ω2 is R ADR The weights, and ω1 + ω2 = 1, R Coverage For coverage, R Penalty This is a penalty item. More specifically, the peak-to-trough ratio and the absorbed dose ratio are standard parameters in this field. The peak-to-trough ratio is a core indicator for evaluating treatment efficacy, reflecting the contrast between high and low dose regions. The absorbed dose ratio is used to evaluate the effective treatment range and calculate the proportion of the high-dose region in the total treatment volume.

[0069] The calculation methods for each parameter are as follows.

[0070] Peak-to-trough ratio (R) in radiotherapy PVDR The calculation method is as follows:

[0071]

[0072] Among them, D peak For the average high dose in the lattice target region, D valley The average low dose is in the extralattice region, and Dpeak and D valley All are determined by optimizing the lattice parameters.

[0073] Absorbed dose ratio R ADR The calculation method is as follows:

[0074]

[0075] Among them, D ablative D represents the total dose within the ablation zone. total The total dose to the tumor region, and D ablative and D total All are determined by optimizing the lattice parameters.

[0076] Coverage R Coverage The calculation method is as follows:

[0077] Wherein, Coverage Area is the volume of the tumor covered by the lattice layout and is determined by optimized lattice parameters, and Tumor Area is the total volume of the tumor and is determined by the tumor geometry.

[0078] The penalty term primarily targets unreasonable lattice distribution characteristics, including both density penalty and overlap penalty. The density penalty, in the form of a quadratic function, constrains the lattice density to remain within a reasonable range. The overlap penalty employs a soft constraint, allowing for a small amount of overlap but imposing a corresponding penalty; this design provides greater optimization space. The penalty term R... Penalty The calculation method is as follows:

[0079] R Penalty =R Penalty-Density +R Penalty-Overlap ;

[0080] Among them, R Penalty-Density R is a density penalty factor used to prevent the lattice layout from being too dense or too sparse. Penalty-Overlap This is an overlap penalty factor used to prevent excessive overlap between lattices from causing uneven dose distribution.

[0081] Furthermore, in, The average coverage volume for each lattice, Optimal Density is the preset optimal lattice density. R Penalty-Overlap =∈·Overlap Area, where ∈ is the constraint factor and Overlap Area is the area of ​​the overlapping region.

[0082] S7. When the reward value does not meet the preset conditions, the value network is used to evaluate the policy network to obtain the state value. The state value is used to guide the policy network to iteratively update the lattice optimization policy. For example... Figure 1 and 2 As shown, the value network is used to evaluate the current optimized state of the lattice, that is, to evaluate the output of the policy network. The input layer of the value network is the same as that of the policy network, ensuring that the value evaluation is based on the same information features, thereby guaranteeing the accuracy of the evaluated state value. The hidden layer is a fully connected layer with 128 nodes, also using the ReLU activation function. The output layer has only one node, which generates the current state value. The state value reflects the quality of the optimized lattice parameters and provides feedback to the policy network on the optimization direction. The core optimization objective of the value network can be expressed as: Where, r t (θ) represents the probability ratio of the new and old strategies, and ∈ is the pruning parameter used to limit the magnitude of strategy updates. In this embodiment, ∈ is set to 0.2.

[0083] In this invention, the reward calculation module uses the peak-to-trough ratio of dose, the absorbed dose ratio, and tumor coverage as the main optimization objectives, and combines these with penalty terms for lattice density and overlap to achieve comprehensive control over dose distribution. In its implementation, the system employs an adaptive weight adjustment mechanism. The weight coefficients are dynamically adjusted based on the changing trends of various indicators during the optimization process. When an indicator deviates significantly from expectations, its weight is increased accordingly; when the indicators tend to balance, the weight remains stable. This adaptive mechanism ensures the stability and reliability of the optimization process.

[0084] The optimization process of this invention adopts an iterative design, with each round of optimization including four main steps: state observation, action selection, execution, and evaluation. An optimization controller coordinates each stage and determines the iteration progress based on convergence conditions. The core iterative formula is:

[0085]

[0086] Where θ represents the optimization parameters, α is the learning rate, and L(θ) is the loss function. An adaptive step size mechanism is introduced during the optimization process, which can dynamically adjust the learning rate based on the optimization effect. When the improvement in effect is not significant after several consecutive rounds of optimization, an early stopping mechanism is triggered to avoid overcomputation.

[0087] Finally, after obtaining the optimal lattice parameters, the dose during radiotherapy can be calculated as a reference for the actual radiotherapy process. The calculation process considers the beam physics and tissue inhomogeneity, and the core calculation model is as follows:

[0088]

[0089] Where K is the dose deposition nucleus, ρ is the tissue density distribution, and Φ is the incident photon flux. This method can accurately reflect the influence of lattice layout on dose distribution.

[0090] In addition, a lookup table can be established based on dose distribution data of commonly used configurations in actual radiotherapy processes, which can improve computational efficiency while ensuring computational accuracy, and also provide a reliable foundation for training lattice optimization models.

[0091] The radiotherapy lattice parameter optimization method of the present invention can generate optimal lattice parameters with better peak-to-valley ratio (PVDR) and absorbed dose ratio (ADR), providing doctors with more targeted and effective parameter guidance for radiotherapy of patients, especially in the treatment of large or complex tumors, and can provide beneficial assistance to the radiotherapy process.

[0092] This invention can take multimodal image data, including CT and MRI images, as input and generate tumor geometric features through DICOM format RTSTRUCT files. It has a wide range of applications, and this data-driven precision placement technology can adapt to tumors of different sizes, shapes, and locations, ensuring that the lattice layout strictly covers the tumor area, thus having higher accuracy and practicality.

[0093] In optimizing lattice parameters, this invention can adaptively adjust the lattice layout parameters according to the specific tumor characteristics of the patient, effectively addressing the dose distribution requirements of irregular and complex-shaped tumors. This overcomes the limitations of fixed parameter settings in existing methods and meets the specific dose distribution needs of different patients during treatment. Based on deep reinforcement learning technology, this invention can automatically generate optimal lattice parameters based on the patient's radiotherapy imaging data, thereby reducing reliance on operator experience and decreasing the time and labor costs of manual setup, thus improving the overall efficiency of the treatment process.

[0094] This invention can support the widespread clinical application of SFRT technology, promote its development in the field of radiation oncology, and improve the success rate of clinical treatment and the quality of life of patients.

[0095] In summary, this invention, by introducing deep reinforcement learning technology, aims to overcome the limitations of existing technologies, improve the accuracy and personalization of spatial segmentation radiotherapy, and ultimately achieve more efficient and safer radiotherapy results.

[0096] Example 2.

[0097] This embodiment is a further optimization based on Embodiment 1. Specifically, Embodiment 2 also includes the following steps.

[0098] After outputting the optimal lattice parameters, empirical data is generated based on the current tumor geometry, initial lattice parameters, and optimal lattice parameters, and added to the experience pool. The empirical data contains complete optimization process information, including parameter change paths, interim evaluation results, and final outcome data. The experience pool is managed using a hierarchical index structure, supporting rapid retrieval based on tumor features. When processing new cases, the system selects similar cases from the experience pool through feature matching, extracts their optimization experience to guide the current parameter optimization process, significantly shortening optimization time and improving optimization effectiveness.

[0099] Furthermore, the method for generating experience includes: First, establishing a first mapping relationship between tumor geometric features and optimal lattice parameters; then, establishing an extended relationship between tumor derived features and optimal lattice parameters, where tumor derived features include centroid location, principal axis direction, spatial anisotropy, tissue density distribution characteristics, spatial distribution of surrounding organs, and risk data of surrounding organs; next, establishing a second mapping relationship between initial lattice parameters and optimal lattice parameters; subsequently, combining tumor geometric features, initial lattice parameters, and optimal lattice parameters into experience based on the first, extended, and second mapping relationships, and adding it to the experience pool; next, clustering the experience based on tumor type, tumor staging results, and tumor geometric features, thereby establishing a connection between the experience and a pre-defined experience index table; finally, optimizing the way the value network evaluates the state value of the strategy network based on the experience pool. It should be noted that the data such as tumor derived features, tumor type, and tumor staging results can all be obtained from the patient's diagnostic results, and will not be elaborated further here.

[0100] When optimizing radiotherapy lattice parameters for other patients, we can first select the experience that is closest to the patient's actual situation from the experience pool based on the experience index table. Then, the first mapping relationship in this experience is used as a guiding parameter and input into the policy network, thereby accelerating the process of generating lattice optimization strategies. On the other hand, the initial lattice parameters are corrected based on the second mapping relationship in this experience. Specifically, a feasible region can be generated based on the second mapping relationship and the optimal lattice parameters. When the initial lattice parameters exceed the feasible region, they are corrected to avoid the initial lattice parameters deviating too much from reality and to further accelerate the overall process of lattice parameter optimization.

[0101] Example 3.

[0102] This embodiment is a further improvement based on Embodiment 2, as detailed below.

[0103] While the second embodiment allows for faster optimization of radiotherapy lattice parameters, each optimization still requires multiple iterations. In practical applications, a new case often requires 100-200 iterations, resulting in a significant overall time consumption. To significantly improve efficiency and shorten optimization time, the third embodiment utilizes a neural network model built based on transfer learning, enabling rapid optimization of radiotherapy lattice parameters even for new cases.

[0104] First, an autoencoder-based feature dimensionality reduction and standardization method is employed to transform experience into feature vectors, ensuring both expressive and computational efficiency. The autoencoder, through a multi-layered neural network structure, compresses high-dimensional features into a low-dimensional latent space during the encoding phase, and then reconstructs the original features during the decoding phase. This process not only achieves feature dimensionality reduction but also preserves key correlations between features. Standardization ensures scale consistency across different dimensions of features, effectively avoiding the impact of dimensional differences on subsequent optimization. This multi-level, multi-dimensional feature representation provides a reliable data foundation for subsequent parameter optimization, effectively supporting knowledge transfer and rapid optimization.

[0105] Second, after each radiotherapy lattice parameter optimization process, i.e., after the optimal lattice parameters are generated, in addition to generating experience, all currently optimized lattice parameters, their corresponding reward values, and state evaluation results are recorded. Particular attention is paid to the impact of parameter changes on dose distribution, including the changing trends of key indicators such as peak-to-trough ratio (PVDR) and target coverage. Furthermore, the decision-making basis of the strategy network and the state value evaluated by the value network are also recorded.

[0106] Third, based on the above data, a typical case library is established, selecting and storing representative typical cases according to different types of tumor characteristics and treatment needs. Each typical case contains a complete optimization history, from the initial lattice parameter setting to the optimal lattice parameter output. The typical case library is classified and managed according to the morphological characteristics, location features, and optimization difficulty of the tumor. In the typical case library, to improve data access efficiency and storage space utilization, compression storage technology can be used to reduce data space while maintaining fast retrieval speed. For intermediate states in the optimization process, a keyframe storage strategy is adopted, retaining only state points with significant changes, and reconstructing the complete optimization trajectory through interpolation methods. In addition, a data indexing mechanism is also needed. To enable efficient retrieval in the typical case library, Locality-Sensitive Hashing (LSH) technology is used to construct an efficient case retrieval mechanism. The LSH algorithm designs a family of feature-sensitive hash functions to map similar cases to the same or adjacent hash buckets. This hash mapping maintains local similarity in the feature space, enabling the rapid location of typical cases with similar characteristics to new cases. To improve the accuracy and robustness of retrieval, a multi-table hashing strategy is adopted, which uses multiple independent hash tables for parallel retrieval, effectively reducing the risk of missed detections due to hash collisions.

[0107] Fourth, a multi-feature weighted similarity calculation method is adopted to facilitate rapid querying of the typical case database. Specifically, when calculating similarity, multiple dimensions such as tumor geometric features, centroid location, and dosage requirements are comprehensively considered. Weighted Euclidean distance or cosine similarity metrics are used to accurately assess the similarity between new cases and cases in the typical case database, thereby quickly retrieving similar cases. To further improve retrieval efficiency, a three-dimensional index structure is established in the typical case database. At the top level, coarse-grained classification is performed based on the basic type and approximate location of the tumor; the middle layer is further subdivided based on more detailed features, such as principal axis direction, spatial anisotropy, and tissue density distribution characteristics; the bottom level contains specific cases and their complete feature information. This hierarchical structure allows for a top-down retrieval strategy, quickly narrowing the search scope and improving retrieval efficiency. Simultaneously, dynamic updates to the index structure are supported, automatically adjusting and optimizing the index structure as new cases are added, maintaining the stability of retrieval performance.

[0108] Fifth, the predicted values ​​of the initial lattice parameters are generated using the nearest neighbor case parameter interpolation method. After retrieving similar cases, instead of simply using the initial or optimal lattice parameters of a single most similar case, multiple highly similar cases are selected for parameter interpolation. The interpolation process employs a distance-weighted strategy, meaning that cases with higher similarity have greater weight in parameter prediction. Simultaneously, a time decay factor is considered, ensuring that the parameters of newer cases have higher reference value, thereby guaranteeing that the prediction results reflect the latest optimization experience. Furthermore, considering that even similar tumors may have differences in local geometric features, a parameter adjustment mechanism based on local features is designed. By analyzing the differences between new cases and reference cases in local regions, including curvature changes and boundary complexity, the initial parameters obtained from interpolation are adjusted in a targeted manner. This adjustment not only considers differences in geometric features but also incorporates the positional constraints of surrounding organs and dose protection requirements, ensuring that the adjusted parameters better meet actual treatment needs.

[0109] Building upon the aforementioned mechanisms, an ensemble learning approach is employed to integrate the results of multiple prediction models to enhance the stability and reliability of parameter prediction. Specifically, this involves training multiple base predictors based on different feature subsets, each focusing on a specific combination of features, such as tumor morphology predictors, location feature predictors, and boundary feature predictors. The outputs of these predictors are then weighted and combined using a random forest (Bagging) strategy and a gradient boosting tree (Boosting) strategy to generate the final parameter prediction result. Furthermore, a prediction uncertainty assessment mechanism is introduced. When the variance of the prediction result exceeds a preset threshold, the system automatically expands the exploration scope and reduces the intensity of local searches to ensure the reliability of the optimized results.

[0110] More specifically, in the pre-training phase of the policy network, historical optimization data is fully utilized to construct the initial policy model. By learning from the optimization trajectories of numerous successful cases, the policy network can grasp the basic rules of parameter adjustment. The pre-training process adopts a hierarchical training strategy, first learning general parameter adjustment patterns, and then performing specialized training for specific types of tumors, forming a hierarchical optimization policy library.

[0111] To facilitate knowledge transfer between different tumor features, a feature adaptation layer was designed. This layer uses feature mapping and distribution alignment techniques to match the feature distribution of typical cases with that of new cases, reducing the impact of inter-domain differences on policy transfer. Simultaneously, a progressive fine-tuning mechanism is employed to gradually adapt to the features of new cases while maintaining the model's basic performance, ensuring the stability and effectiveness of the optimization strategy.

[0112] In terms of value assessment transfer, a benchmark model for value assessment was constructed by analyzing the correlation between parameter adjustments and treatment effects in typical cases. This model not only transfers specific assessment values ​​but also assessment criteria and weight allocation schemes to ensure the comparability and consistency of assessment results. The reliability of the value assessment is guaranteed through a confidence level assessment mechanism. The uncertainty of each assessment result is calculated, and the confidence level is determined based on factors such as feature similarity and sample size. For assessment results with low confidence levels, the proportion of exploratory searches is increased to avoid over-reliance on uncertain experience. The transfer weights are dynamically adjusted based on the confidence level of the assessment and the actual optimization effect, achieving adaptive control of experience transfer.

[0113] The search space guidance mechanism significantly improves parameter optimization efficiency by leveraging historical optimization experience. Analyzing the parameter distribution patterns in typical cases determines the approximate range of effective parameters, thus narrowing the search space. This narrowing is not a simple range restriction, but a dynamic adjustment based on the characteristics of the current cases, ensuring the rationality of the search space. An adaptive parameter sampling strategy is employed, guiding the sampling process based on the probability density of parameter distribution in historical experience. Regarding the balance between exploration and utilization, an improved UCB (Upper Confidence Bound) algorithm is used to dynamically adjust the ratio of exploring new parameter combinations to utilizing known excellent parameters, avoiding getting trapped in local optima.

[0114] The convergence acceleration mechanism improves optimization efficiency through multiple strategies. It utilizes historical optimization trajectory data to predict the optimal direction for parameter adjustment, reducing ineffective parameter exploration. During optimization, the learning rate is dynamically adjusted based on parameter convergence. A larger learning rate is used in the early stages to quickly approach the optimal solution region, while the learning rate is reduced as convergence approaches for fine-tuning. To avoid over-iteration, an intelligent early stopping mechanism is designed. By monitoring the changing trends of optimization metrics, the optimization process automatically terminates when the improvement after multiple consecutive optimization rounds falls below a preset threshold. Simultaneously, the optimal parameter combination during the optimization process is recorded, allowing for a rollback to the previous optimal state when performance degradation occurs, ensuring the reliability of the optimization results.

[0115] like Figure 4 As shown, the present invention further provides a radiotherapy lattice parameter optimization system based on spatial segmentation technology, including an input module, a tumor identification module, a lattice parameter optimization module, a lattice arrangement and DICOM update module, and an output module.

[0116] The input module is used to acquire and preprocess the patient's radiotherapy image data. Specifically, the input module includes a file loading unit and a parameter input unit. The file loading unit is used to read RTSTRUCT files, including those in DICOM format, and the parameter input unit is used to acquire initial parameters provided by the user, such as lattice spacing and radius. The lattice parameter optimization model can continuously optimize based on these input initial parameters to obtain the optimal lattice parameters.

[0117] The tumor identification module is used to calculate the geometric features of the tumor based on radiotherapy image data and preset geometric constraints. Specifically, the tumor identification module includes a DICOM parsing unit and a GTV contour extraction unit. The DICOM parsing unit is used to parse the read DICOM format RTSTRUCT file, and the GTV contour extraction unit is used to extract the tumor geometric features through contour extraction based on the parsing results of the DICOM parsing unit.

[0118] The lattice parameter optimization module is used to identify tumor geometric features using a lattice parameter optimization model to obtain optimal lattice parameters. Specifically, the lattice parameter optimization module includes an agent initialization unit, a lattice arrangement unit, a reward calculation unit, and a policy update unit. The agent initialization unit initializes the PPO algorithm; the lattice arrangement unit performs lattice arrangement based on the initial parameters and continuously optimizes the lattice arrangement using the PPO algorithm to obtain the original lattice parameters; the reward calculation unit determines whether the original lattice parameters satisfy the PVDR and ADR conditions to determine the optimal lattice parameters; and the policy update unit updates and optimizes the policy network and value network.

[0119] The lattice arrangement and DICOM update module is used to generate lattice arrangement data based on optimal lattice parameters and to generate a data transfer file. Specifically, the lattice arrangement and DICOM update module includes a lattice arrangement unit and a DICOM update unit. The lattice arrangement unit is used to perform lattice arrangement based on optimal lattice parameters and generate optimal lattice point distribution data; the DICOM update unit is used to write the optimal lattice point distribution data into a DICOM format RTSTRUCT file.

[0120] The output module includes a file saving unit and a result evaluation unit. The file saving unit is used to save the generated RTSTRUCT file to the storage medium, and the result evaluation unit is used to evaluate the optimal lattice point distribution data and output an evaluation report.

[0121] When this system is used to implement the above-described Embodiment 2, it further includes a feature association analysis unit, a case retrieval unit, a strategy optimization unit, and a parameter prediction unit.

[0122] The feature correlation analysis unit is used to establish the mapping relationship between tumor features and optimization parameters. By extracting basic feature information such as tumor volume, shape, and boundary, and combining it with environmental features such as the distribution of surrounding organs, the unit constructs a complete feature vector. The correspondence between these feature information and the optimal lattice parameters is systematically stored and managed, providing a data foundation for subsequent optimization processes.

[0123] The case retrieval unit identifies cases with similar characteristics from the experience pool. By establishing a multi-level index structure, the system can efficiently calculate the feature similarity between cases, achieving rapid retrieval and matching. This similarity-based retrieval mechanism effectively reduces the parameter search space, providing valuable reference information for the optimization process.

[0124] The strategy optimization unit dynamically adjusts the optimization strategy based on historical optimization experience. This unit optimizes the training process of the strategy network by analyzing historical data and adjusting the reference criteria for value evaluation. This strategy update mechanism enables the optimization process to better adapt to different types of tumor characteristics, improving optimization efficiency.

[0125] The parameter prediction unit is responsible for generating the initial parameters for optimization. This unit predicts the feasible region of the parameters by analyzing historical data, evaluates the rationality of parameter combinations, and can dynamically adjust the parameter generation strategy. This parameter prediction mechanism can significantly improve the quality of initial parameters and accelerate the optimization convergence speed.

[0126] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or modules may be electrical, mechanical, or other forms.

[0127] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0128] The present invention also provides a computer device, including a memory and a processor.

[0129] Memory is used to store computer programs.

[0130] A processor is used to read and execute the computer program to implement the above-described method for optimizing radiotherapy lattice parameters based on spatial segmentation technology.

[0131] Finally, the present invention provides a storage medium storing a computer program that, when executed, implements the above-described method for optimizing radiotherapy lattice parameters based on spatial segmentation technology.

[0132] Based on this storage medium, the technical solution of this disclosure, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0134] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for optimizing radiotherapy lattice parameters based on spatial segmentation technology, characterized in that, Includes the following steps: Acquire the patient's radiotherapy imaging data; The geometric features of the tumor are calculated based on radiotherapy imaging data and pre-set geometric constraints. The tumor geometric features and initial lattice parameters are input into a lattice parameter optimization model constructed based on a proximal policy optimization algorithm. The lattice parameter optimization model includes a policy network, a parameter optimization module, a reward calculation module, and a value network. A policy network is used to iteratively generate a lattice optimization strategy based on tumor geometric features and initial lattice parameters. The parameter optimization module generates optimized lattice parameters based on a lattice optimization strategy. The reward calculation module is used to calculate the reward value of the optimized lattice parameters, and when the reward value meets the preset conditions, the optimized lattice parameters are output as the optimal lattice parameters. The specific methods for calculating rewards include: R=ω1·R PVDR +ω2·R ADR +ω3·R Coverage -ω4·R Penalty ; Where R is the total reward value, R PVDR R is the peak-to-trough ratio of the dose in radiotherapy. ADR ω1 is the absorbed dose ratio, where R is the dose ratio. PVDR The weights, ω2 is R ADR The weights, and ω1 + ω2 = 1, R Coverage For coverage, R Penalty This is a penalty item; Coverage R Coverage The calculation method is as follows: Wherein, Coverage Area is the volume of the tumor covered by the lattice layout and is determined by optimized lattice parameters, and TumorArea is the total volume of the tumor and is determined by the tumor geometry. Penalty item R Penalty The calculation method is as follows: R Penalty =R Penalty-Density +R Penalty-Overlap ; Among them, R Penalty-Density R is the density penalty factor. Penalty-Overlap This is an overlap penalty factor; When the reward value does not meet the preset conditions, the value network is used to evaluate the policy network to obtain the state value. The state value is used to guide the policy network to update the lattice optimization policy through iteration.

2. The method for optimizing radiotherapy lattice parameters based on spatial segmentation technology as described in claim 1, characterized in that, Tumor geometry features include tumor volume, shape factor, and boundary features, where the shape factor is used to describe the tumor as a regular geometry based on radiotherapy imaging data.

3. The method for optimizing radiotherapy lattice parameters based on spatial segmentation technology as described in claim 1, characterized in that, After calculating the geometric features of the tumor, the state space and action space of the lattice parameter optimization model are determined based on the geometric features of the tumor.

4. The method for optimizing radiotherapy lattice parameters based on spatial segmentation technology as described in claim 3, characterized in that, The state-space representation of the lattice parameter optimization model is as follows: S={Volume, Shape Factor, Boundary Features, Current Parameters}; Where Volume is the tumor volume, Shape Factor is the shape factor, Boundary Features are the boundary features, and Current Parameters are the current lattice parameter state; The action space of the lattice parameter optimization model is represented as: A=[Spacing, Radius, Edge distance]; Where Spacing is the distance between lattice points; Radius is the effective radius of the lattice points; and Edge distance is the distance between the lattice and the edge of the tumor target area.

5. The method for optimizing radiotherapy lattice parameters based on spatial segmentation technology as described in claim 1, characterized in that, Peak-to-trough ratio (R) in radiotherapy PVDR The calculation method is as follows: Among them, D peak For the average high dose in the lattice target region, D valley The average low dose is in the extralattice region, and D peak and D valley All are determined by optimizing lattice parameters; Absorbed dose ratio R ADR The calculation method is as follows: Among them, D ablative D represents the total dose within the ablation zone. total The total dose to the tumor region, and D ablative and D total All are determined by optimizing the lattice parameters.

6. A radiotherapy lattice parameter optimization system based on spatial segmentation technology, characterized in that, The system for implementing the radiotherapy lattice parameter optimization method based on spatial segmentation technology as described in claim 1 includes: The input module is used to acquire the patient's radiotherapy image data and to preprocess the radiotherapy image data; The tumor identification module is used to calculate the geometric features of the tumor based on radiotherapy image data and preset geometric constraints. The lattice parameter optimization module is used to identify the optimal lattice parameters by utilizing the lattice parameter optimization model to identify the geometric features of the tumor.

7. The radiotherapy lattice parameter optimization system based on spatial segmentation technology as described in claim 6, characterized in that, The system also includes a lattice arrangement and DICOM update module, which generates lattice arrangement data based on optimal lattice parameters and generates a data transfer file.

Images