A tumor radiotherapy anomaly warning system based on clustering algorithm

Abstract

The present invention relates to the technical field of radiotherapy data analysis, and proposes a tumor radiotherapy anomaly warning system based on a clustering algorithm, aiming to improve the anomaly detection ability and warning accuracy of patients' health data during radiotherapy; by introducing the local outlier factor and adaptive bandwidth optimization technology, the system optimizes the density weighted calculation of the MeanShift clustering algorithm, can accurately identify the internal patterns in the data of tumor patients, and improves the accuracy of preliminary clustering; in the enhanced clustering process, the entropy regularization is used to optimize the P3D clipping method, further optimize the policy network update of the PPO reinforcement learning algorithm, dynamically adjust the clustering features, and enhance the adaptability and accuracy of the system in a dynamic environment; by combining reinforcement learning and clustering analysis technologies, the system provides high-precision and high-real-time support for anomaly warning during radiotherapy, significantly improving the intelligent level and reliability of patient monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of radiotherapy data analysis, and in particular to a tumor radiotherapy anomaly warning system based on a clustering algorithm. Background Art

[0002] With the development of technology, more and more medical fields have started to adopt big data analysis and artificial intelligence technologies to improve efficiency and data management levels. However, traditional anomaly warning systems still have many limitations. First of all, traditional systems often ignore the individual differences of patients and the complexity and diversity of data during radiotherapy, resulting in insufficient sensitivity of the system to abnormal events or the inability to identify potential risks in a timely and effective manner. Secondly, traditional systems have limitations in data analysis capabilities and cannot make full use of multi-dimensional data such as patients' physiological indicators, radiotherapy data, and tumor characteristics for in-depth analysis and pattern recognition, making the warning effect of the system inaccurate and difficult to cope with the changing patient conditions and dynamic changes during radiotherapy, thus reducing the overall performance and reliability of the system. Therefore, there is an urgent need for a tumor radiotherapy anomaly warning system that can more effectively integrate multi-dimensional data, improve the accuracy of anomaly detection, and dynamically optimize clustering results to address the problems of poor accuracy and low performance in the prior art and enhance the intelligence and real-time nature of patient monitoring during radiotherapy. Summary of the Invention

[0003] The present invention proposes a tumor radiotherapy anomaly warning system based on a clustering algorithm, aiming to improve the accuracy of anomaly detection and the performance of the system. First of all, in the initial clustering process, the system adopts the Local Outlier Factor (LOF) and adaptive bandwidth optimization technology to optimize the density-weighted calculation of the MeanShift clustering algorithm. By introducing LOF to detect the abnormality of data points and combining the adaptive bandwidth to adjust the bandwidth parameter of the clustering algorithm, the system can accurately identify the internal patterns in complex patient data, thereby generating more effective initial clustering features. Secondly, in the enhanced clustering process, entropy regularization is introduced to optimize the P3D pruning method to optimize the policy network update of the Proximal Policy Optimization (PPO) reinforcement learning algorithm. Entropy regularization helps to maintain the balance between exploration and exploitation, while the P3D pruning method limits the amplitude of policy updates, enhancing the stability and convergence of the clustering results. Finally, the clustered data will be subjected to anomaly detection through a Generative Adversarial Network (GAN). The generative adversarial network can identify potential abnormal situations based on the clustered feature data and further optimize the detection effect. Through this series of technical optimizations, the present invention can dynamically adjust the clustered feature data, provide high-precision and high-stability feature data support, thereby enhancing the system's early warning ability and overall performance for anomalies during tumor radiotherapy.

[0004] The present invention provides a tumor radiotherapy anomaly warning system based on a clustering algorithm. The system includes a data acquisition module, a data preprocessing module, a preliminary clustering module, an enhanced clustering module, an anomaly detection module, a warning decision-making module, and a visualization module;

[0005] The data acquisition module collects relevant data during tumor radiotherapy, including patient physiological indicators, radiotherapy treatment data, radiotherapy reaction data, and tumor feature data, and integrates them to obtain comprehensive patient health data;

[0006] The data preprocessing module performs data cleaning, normalization, feature selection, dimensionality reduction, and data transformation on the comprehensive patient health data to generate preprocessed comprehensive patient health data;

[0007] The preliminary clustering module introduces the local outlier factor and an adaptive bandwidth to optimize the update mechanism of the MeanShift clustering algorithm, constructs an enhanced MeanShift clustering algorithm update mechanism, optimizes the MeanShift clustering algorithm through the enhanced MeanShift clustering algorithm update mechanism, constructs a strengthened MeanShift clustering algorithm, and extracts the preliminary clustering features of the preprocessed comprehensive patient health data through the strengthened MeanShift clustering algorithm to generate preliminary clustering feature data;

[0008] The enhanced clustering module introduces the PPO algorithm and the P3D clipping method, optimizes the P3D clipping method using entropy regularization, constructs an entropy-P3D clipping method, optimizes the policy network update of the PPO algorithm through the entropy-P3D clipping method, constructs an entropy-P3D-PPO algorithm, and dynamically adjusts the preliminary clustering feature data through the entropy-P3D-PPO algorithm to generate optimized clustering feature data;

[0009] The anomaly detection module establishes a generative adversarial network, inputs the preprocessed comprehensive patient health data, the preliminary clustering feature data, and the optimized clustering feature data into the generative adversarial network for anomaly detection, and generates anomaly detection results;

[0010] The warning decision-making module generates a radiotherapy anomaly warning signal based on the anomaly detection results to remind medical staff to take timely treatment measures;

[0011] The result visualization module visualizes the anomaly detection results through charts to generate visualization charts. Through the comprehensive patient health data and the visualization charts, it helps medical staff intuitively understand the patient's radiotherapy status and anomaly risks;

[0012] The hierarchical structure of the entropy-P3D-PPO algorithm includes an initialization unit, an action selection unit, a reward calculation unit, a reward weighting unit, a clustering data adjustment unit, a policy clipping optimization unit, and an iterative optimization unit.

[0013] Further, the process of the preliminary clustering module generating preliminary clustering feature data specifically includes the following:

[0014] Step S1: Data partitioning: The preprocessed comprehensive patient health data is evenly partitioned into multiple data blocks to obtain partitioned patient health data, making the preprocessed comprehensive patient health data have high internal similarity and facilitating clustering analysis;

[0015] Step S2: Initialize MeanShift clustering: Initialize the partitioned patient health data to generate partitioned patient data blocks. Select data points from the partitioned patient data blocks as the current clustering centers, and set the window size and Gaussian kernel function of the MeanShift clustering algorithm;

[0016] Step S3: Clustering iteration: Update the current clustering centers iteratively through the enhanced MeanShift clustering algorithm update mechanism to generate convergent clustering centers;

[0017] Step S4: Determine clustering labels: According to the convergent clustering centers, assign the closest clustering center to each data point in the partitioned patient data blocks, identify the clustering cluster to which each data point belongs, and generate clustering labels;

[0018] Step S5: Generate data: According to the convergent clustering centers and clustering labels, extract clustering feature data to generate preliminary clustering feature data. The preliminary clustering feature data includes the clustering center position, clustering size, clustering density, clustering dispersion, average eigenvalue of the cluster, similarity between clusters, variability between clusters, and clustering label distribution.

[0019] Further, Step S3 specifically includes the following steps:

[0020] Step S31: Density weighted calculation: Introduce the local outlier factor to perform outlier detection on each data point in the partitioned patient data blocks to generate LOF outlier detection values; Combine the LOF outlier detection values, calculate the density weighting of each data point in the partitioned patient data blocks through the Gaussian kernel function, introduce an adaptive bandwidth to optimize the window size, and control the bandwidth parameter of the Gaussian kernel function through the window size to generate weighted density values. The formula used is as follows:

[0021] ;

[0022] where, represents the index of the data point in the partitioned patient data blocks, represents the current clustering center, represents the data point, represents the data point for the weighted density value of the current clustering center , represents the data point The square of the Euclidean distance to the current cluster center is denoted as the square of the adaptive bandwidth, is denoted as the adjustment factor,

[0023] Step S32: Update the cluster center: According to the weighted density value, update the current cluster center in combination with the MeanShift update mechanism, and calculate the change amount of the current cluster center. The formula used is as follows:

[0024] ;

[0025] wherein, is denoted as the iteration index, is denoted as the cluster center at the -th iteration, is denoted as the weighted density value of the data point with respect to denoted as the number of data points in the current cluster;

[0026] Step S33: Iterative convergence: Set the iteration threshold, and iterate Steps S31 - S32 until the change amount of the current cluster center is less than the iteration threshold to generate a converged cluster center.

[0027] Furthermore, the initialization unit represents the preliminary clustering feature data as the state information in reinforcement learning to generate a clustering state representation. The clustering state representation includes the central position of the cluster, the average size of the object, and the number of clusters;

[0028] The action selection unit selects an action according to the clustering state representation through the current policy network as the decision-making action. The decision-making actions include maintaining the cluster, merging the clusters, and splitting the clusters;

[0029] The reward calculation unit calculates the clustering compactness reward value, the object consistency reward value, the cluster number control reward value, and the cluster separation reward value of the decision-making action through the clustering compactness, the object area variance, the number of clusters, and the distance between the clusters;

[0030] The reward weighting unit weights the clustering compactness reward value, the object consistency reward value, the cluster number control reward value, and the cluster separation reward value to obtain a weighted reward value;

[0031] The clustering data adjustment unit evaluates the effect of the decision-making action according to the weighted reward value, updates the current policy network, and adjusts the clustering state of the preliminary clustering feature data.

[0032] Further, the policy pruning and optimization unit prunes the current policy network through the entropy-P3D pruning method to limit the update amplitude of the current policy network. The formula used is as follows:

[0033] ;

[0034] Wherein, represents the action, represents the state, represents the policy, represents the entropy of the policy, represents the policy selects the action under the state with probability, represents the logarithm of the policy probability;

[0035] ;

[0036] Wherein, represents the time step, represents the action at the -th time step, represents the adjusted reward, represents the original reward, represents the entropy regularization coefficient;

[0037] ;

[0038] Wherein, represents the probability of the new policy selecting the action, represents the probability of the old policy selecting the action, represents the pruning threshold, represents the difference between the new and old policies, represents pruning the policy difference.

[0039] Further, the iterative optimization unit sets an iterative threshold and performs multiple iterative optimizations by repeatedly executing the action selection unit, the reward calculation unit, the reward weighting unit, the clustering data adjustment unit, and the policy pruning and optimization unit until the iterative threshold is reached, generating optimized clustering feature data.

[0040] Adopting the above solution, the beneficial effects obtained by the present invention are as follows:

[0041] By introducing the Local Outlier Factor (LOF) and adaptive bandwidth optimization technology, the present invention optimizes the density-weighted calculation of the MeanShift clustering algorithm, significantly enhancing the system's ability to process patients' health data. Through this technical improvement, the system can more accurately identify the inherent patterns in the data. Especially when faced with complex and diverse patient data, it can perform precise clustering based on individual differences. This not only improves the accuracy of the initial clustering but also provides more reliable data support for subsequent anomaly detection and early warning decisions, solving the limitations and insufficient sensitivity problems of traditional systems in processing multi-dimensional data.

[0042] During the enhanced clustering process, the present invention introduces an entropy regularization optimization P3D clipping method to optimize the policy network update of the PPO reinforcement learning algorithm. This technology effectively enhances the system's adaptability in a dynamically changing environment, enabling the system to maintain stability and efficiency under changing radiotherapy data and patients' health statuses. By optimizing the clustering process, the system can dynamically adjust clustering features according to real-time data, avoiding common problems such as overfitting and insufficient accuracy in traditional systems, and greatly improving the accuracy and real-time performance of early warnings, thus enabling the system to more reliably identify potential abnormal events.

[0043] Finally, the present invention performs anomaly detection on the clustered data through a Generative Adversarial Network (GAN), deeply mines the multi-level clustering features, effectively identifies abnormal patterns in the data, provides more intuitive and reliable early warning information for medical staff, helps to identify potential risks in advance, improves the intelligent level of the system in tumor radiotherapy monitoring, and ensures the safety of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the modules of a tumor radiotherapy anomaly early warning system based on a clustering algorithm proposed by the present invention;

[0045] Figure 2 It is a schematic diagram of the process of generating preliminary clustering feature data by the preliminary clustering module in Embodiment 2;

[0046] Figure 3 It is a schematic diagram of the process of the updated mechanism of the enhanced MeanShift clustering algorithm in Embodiment 3. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] Example 1. According to Figure 1 , the present invention provides a tumor radiotherapy anomaly warning system based on a clustering algorithm. The system includes a data acquisition module, a data preprocessing module, a preliminary clustering module, an enhanced clustering module, an anomaly detection module, a warning decision-making module, and a visualization module;

[0049] The data acquisition module collects relevant data during tumor radiotherapy, including patient physiological indicators, radiotherapy treatment data, radiotherapy reaction data, and tumor characteristic data, and integrates them to obtain comprehensive patient health data;

[0050] The data preprocessing module performs data cleaning, normalization, feature selection, dimensionality reduction, and data conversion on the comprehensive patient health data to generate preprocessed comprehensive patient health data;

[0051] The preliminary clustering module introduces the local outlier factor and an adaptive bandwidth to optimize the update mechanism of the MeanShift clustering algorithm, constructs an enhanced MeanShift clustering algorithm update mechanism, optimizes the MeanShift clustering algorithm through the enhanced MeanShift clustering algorithm update mechanism, constructs a strengthened MeanShift clustering algorithm, and extracts the preliminary clustering features of the preprocessed comprehensive patient health data through the strengthened MeanShift clustering algorithm to generate preliminary clustering feature data;

[0052] The enhanced clustering module introduces the PPO algorithm and the P3D clipping method, optimizes the P3D clipping method using entropy regularization, constructs an entropy - P3D clipping method, optimizes the policy network update of the PPO algorithm through the entropy - P3D clipping method, constructs an entropy - P3D - PPO algorithm, and dynamically adjusts the preliminary clustering feature data through the entropy - P3D - PPO algorithm to generate optimized clustering feature data;

[0053] The anomaly detection module establishes a generative adversarial network, inputs the preprocessed comprehensive patient health data, the preliminary clustering feature data, and the optimized clustering feature data into the generative adversarial network for anomaly detection, and generates anomaly detection results;

[0054] The warning decision-making module generates a radiotherapy anomaly warning signal based on the anomaly detection results to remind medical staff to take timely treatment measures;

[0055] The result visualization module visualizes the anomaly detection results through charts to generate visualization charts, and helps medical staff intuitively understand the patient's radiotherapy status and anomaly risks through the comprehensive patient health data and the visualization charts;

[0056] The hierarchical structure of the entropy - P3D - PPO algorithm includes an initialization unit, an action selection unit, a reward calculation unit, a reward weighting unit, a clustering data adjustment unit, a policy clipping optimization unit, and an iterative optimization unit.

[0057] Example 2. According to Figure 2 , this example is based on Example 1. In this example, the process of the preliminary clustering module generating preliminary clustering feature data specifically includes the following:

[0058] Step S1: Data partitioning: The preprocessed comprehensive patient health data is evenly partitioned into multiple data blocks to obtain partitioned patient health data, making the preprocessed comprehensive patient health data have high internal similarity and facilitating clustering analysis;

[0059] The number of evenly partitioned data blocks is: 90;

[0060] Step S2: Initialize MeanShift clustering: Initialize the partitioned patient health data to generate partitioned patient data blocks, select a data point from the partitioned patient data blocks as the current clustering center, and set the window size and Gaussian kernel function of the MeanShift clustering algorithm;

[0061] Step S3: Clustering iteration: Update the current clustering center iteratively through the enhanced MeanShift clustering algorithm update mechanism to generate a convergent clustering center;

[0062] Step S4: Determine the clustering label: According to the convergent clustering center, assign the nearest clustering center to each data point in the partitioned patient data blocks, identify the clustering cluster to which each data point belongs, and generate a clustering label;

[0063] Step S5: Generate data: According to the convergent clustering center and the clustering label, extract clustering feature data to generate preliminary clustering feature data. The preliminary clustering feature data includes the clustering center position, clustering size, clustering density, clustering dispersion, average feature value of the clustering, similarity between clusters, variability between clusters, and clustering label distribution.

[0064] Example 3. According to Figure 3 , this example is based on Example 2. In this example, Step S3 specifically includes the following steps:

[0065] Step S31: Density weighted calculation: Introduce the local outlier factor to perform outlier detection on each data point in the partitioned patient data blocks to generate LOF outlier detection values; Combine the LOF outlier detection values, calculate the density weighting of each data point in the partitioned patient data blocks through the Gaussian kernel function, introduce an adaptive bandwidth to optimize the window size, and control the bandwidth parameter of the Gaussian kernel function through the window size to generate a weighted density value. The formula used is as follows:

[0066] ;

[0067] Among them, Indicates the index of the data point in the partitioned patient data block, Indicates the current cluster center, Indicates the data point, Indicates the data point For the current cluster center The weighted density value of, Indicates the data point For the current cluster center The square of the Euclidean distance of, Indicates the square of the adaptive bandwidth, Indicates the adjustment factor, Indicates the local outlier factor of the data point, used to measure the abnormality degree of the data point;

[0068] Step S32: Update the cluster center: According to the weighted density value, update the current cluster center by combining the MeanShift update mechanism, and calculate the change amount of the current cluster center. The formula used is as follows:

[0069] ;

[0070] Among them, Indicates the iteration index, Indicates The cluster center at the Indicates The cluster center at the Indicates the data point To The weighted density value of, Indicates the number of data points in the current cluster;

[0071] Step S33: Iterative convergence: Set the iteration threshold, and iterate steps S31 - S32 until the change amount of the current cluster center is less than the iteration threshold, and generate the convergent cluster center.

[0072] Example 4, this example is based on Example 2. In this example, step S3 specifically includes the following steps:

[0073] Step Q1: Density weighted calculation: Calculate the density weighting of each data point in the partitioned patient data block through the Gaussian kernel function, and control the bandwidth parameter of the Gaussian kernel function through the window size to generate the weighted density value;

[0074] Step Q2: Update the cluster center: According to the weighted density value, update the current cluster center by combining the MeanShift update mechanism, and calculate the change amount of the current cluster center. The formula used is as follows:

[0075] ;

[0076] Among them, represents the iteration index, represents the cluster center at the [[ID=]], represents the cluster center at the [[ID=]] represents the data point for the weighted density value of represents the number of data points in the current cluster;

[0077] Step Q3: Iterative convergence: Set the iteration threshold, and perform iterative steps S31 - S32 until the change in the current cluster center is less than the iteration threshold, generating a converged cluster center; the threshold is 0.0001.

[0078] Example 5, this example is based on Example 3. In this example, the initialization unit represents the preliminary clustering feature data as the state information in reinforcement learning, generating a clustering state representation. The clustering state representation includes the center position of the cluster, the average size of the object, and the number of clusters;

[0079] The action selection unit selects an action according to the clustering state representation through the current policy network as the decision-making action. The decision-making action includes maintaining the cluster, merging the clusters, and splitting the clusters;

[0080] The reward calculation unit calculates the cluster compactness reward value, object consistency reward value, cluster number control reward value, and cluster separation reward value of the decision-making action through the cluster compactness, object area variance, number of clusters, and distance between clusters;

[0081] The reward weighting unit weights the cluster compactness reward value, object consistency reward value, cluster number control reward value, and cluster separation reward value to obtain a weighted reward value;

[0082] The clustering data adjustment unit evaluates the effect of the decision-making action according to the weighted reward value, updates the current policy network, and adjusts the clustering state of the preliminary clustering feature data.

[0083] Example 6, this example is based on Example 5. In this example, the policy pruning and optimization unit prunes the current policy network through the entropy - P3D pruning method to limit the update amplitude of the current policy network. The formula used is as follows:

[0084] ;

[0085] Where represents the action, represents the state, represents the policy, represents the entropy of the policy, represents the policy The probability of selecting an action in the state , represents the logarithm of the policy probability;

[0086] ;

[0087] wherein, represents the time step, represents the th action at the time step, represents the adjusted reward, represents the original reward, represents the entropy regularization coefficient;

[0088] ;

[0089] wherein, represents the probability of the new policy selecting an action, represents the probability of the old policy selecting an action, represents the clipping threshold, represents the difference between the new and old policies, represents clipping the policy difference.

[0090] Example 7: This example is based on Example 5. In this example, the policy clipping and optimization unit clips the current policy network through the P3D clipping method to limit the update amplitude of the current policy network. The formula used is as follows:

[0091] ;

[0092] wherein, represents the probability of the new policy selecting an action, represents the probability of the old policy selecting an action, represents the clipping threshold, represents the difference between the new and old policies, represents clipping the policy difference.

[0093] Example 8: This example is based on Example 6. In this example, the iterative optimization unit sets the maximum number of iterations and performs multiple iterative optimizations by repeatedly executing the action selection unit, the reward calculation unit, the reward weighting unit, the clustering data adjustment unit, and the policy clipping and optimization unit until the iteration threshold is reached to generate optimized clustering feature data;

[0094] Perform 60 iterative optimizations.

[0095] The above description of the present invention and its embodiments is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural forms and embodiments to this technical solution without creative work and without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A tumor radiotherapy abnormality early warning system based on a clustering algorithm, comprising a data preprocessing module, wherein the data preprocessing module generates preprocessed comprehensive patient health data; characterized in that: The system also includes a preliminary clustering module and an enhanced clustering module; The preliminary clustering module introduces a local outlier factor and an adaptive bandwidth to optimize the update mechanism of the MeanShift clustering algorithm, constructs an enhanced MeanShift clustering algorithm update mechanism, optimizes the MeanShift clustering algorithm through the enhanced MeanShift clustering algorithm update mechanism, constructs an enhanced MeanShift clustering algorithm, extracts preliminary clustering features of preprocessed comprehensive patient health data through the enhanced MeanShift clustering algorithm, and generates preliminary clustering feature data; The enhanced clustering module introduces the PPO algorithm and the P3D clipping method, uses entropy regularization to optimize the P3D clipping method, constructs the entropy-P3D clipping method, optimizes the policy network update of the PPO algorithm through the entropy-P3D clipping method, constructs the entropy-P3D-PPO algorithm, dynamically adjusts the preliminary clustering feature data through the entropy-P3D-PPO algorithm, and generates optimized clustering feature data; The anomaly detection module establishes a generative adversarial network, inputs the preprocessed comprehensive patient health data, preliminary clustering feature data and optimized clustering feature data into the generative adversarial network for anomaly detection, and generates anomaly detection results.

2. The tumor radiotherapy abnormality early warning system based on clustering algorithm according to claim 1, characterized in that: The hierarchical structure of the Entropy-P3D-PPO algorithm includes an initialization unit, an action selection unit, a reward calculation unit, a reward weighting unit, a clustering data adjustment unit, a strategy clipping optimization unit, and an iterative optimization unit.

3. The tumor radiotherapy abnormality early warning system based on clustering algorithm according to claim 1, characterized in that: The process of generating preliminary clustering feature data by the preliminary clustering module specifically includes the following contents: Step S1: uniformly partition the pre-processed comprehensive patient health data to obtain partitioned patient health data; Step S2: Initialize the partitioned patient health data, generate partitioned patient data blocks, select data points from the partitioned patient data blocks as current cluster centers, and set the window size and Gaussian kernel function of the MeanShift clustering algorithm; Step S3: Iteratively update the current cluster center through the enhanced MeanShift clustering algorithm update mechanism to generate a convergent cluster center; Step S4: according to the converged cluster center, assign the nearest cluster center to each data point in the partitioned patient data block and generate a cluster label; Step S5: Generate preliminary cluster feature data based on the converged cluster centers and cluster labels.

4. The tumor radiotherapy abnormality early warning system based on clustering algorithm according to claim 3 is characterized by: Step S3 specifically includes the following steps: Step S31: introduce a local outlier factor to perform anomaly detection on each data point in the partitioned patient data block to generate a LOF anomaly detection value; combine the LOF anomaly detection value, calculate the density weight of each data point in the partitioned patient data block through a Gaussian kernel function, introduce an adaptive bandwidth to optimize the window size, and control the bandwidth parameter of the Gaussian kernel function through the window size to generate a weighted density value; Step S32: updating the current cluster center according to the weighted density value in combination with the MeanShift update mechanism, and calculating the change amount of the current cluster center; Step S33: setting an iteration threshold, iterating steps S31 to S32 until the change in the current cluster center is less than the iteration threshold, and generating a converged cluster center.

5. The tumor radiotherapy abnormality early warning system based on clustering algorithm according to claim 2, characterized in that: The initialization unit represents the preliminary clustering feature data as state information to generate a clustering state representation; The action selection unit selects an action as a decision action through the current policy network according to the cluster state representation; The reward calculation unit calculates the cluster density reward value, object consistency reward value, cluster number control reward value and cluster separation reward value of the decision action through cluster density, object area variance, cluster number and distance between clusters; The reward weighting unit weights the cluster compactness reward value, the object consistency reward value, the cluster quantity control reward value and the cluster separation reward value to obtain a weighted reward value; The clustering data adjustment unit evaluates the effect of the decision action according to the weighted reward value, updates the current policy network, and adjusts the clustering state of the preliminary clustering feature data.

6. The tumor radiotherapy abnormality early warning system based on clustering algorithm according to claim 5, characterized in that: The policy pruning optimization unit prunes the current policy network through the entropy-P3D pruning method to limit the update range of the current policy network.

7. The tumor radiotherapy abnormality early warning system based on clustering algorithm according to claim 6, characterized in that: The iterative optimization unit sets a maximum number of iterations, and generates optimized cluster feature data by repeatedly executing the action selection unit, the reward calculation unit, the reward weighting unit, the cluster data adjustment unit, and the strategy clipping optimization unit until an iteration threshold is reached.

Images