VMAT plan patient dose verification prediction method and system based on Bayesian model uncertainty quantification

Abstract

The invention provides a VMAT plan patient dose verification prediction method and system based on Bayesian model uncertainty quantification, and is applied to the technical field of medical data processing. According to the method, a multi-modal classification model is constructed, a radiation field complexity feature and a flux graph are used as input, one-dimensional and two-dimensional branches are used for feature extraction and then fusion, a full connection layer is used for output, an Adam optimizer is used for training, variational Monte Carlo is combined, and data, the model and total uncertainty are calculated through sampling. Based on classification model output, through LASSO dimensionality reduction, in combination with classification probability and uncertainty, MLP prediction GPR is constructed, self-defined weighted MAE training is used, test samples are classified according to clinical thresholds, intervention is triggered by exceeding thresholds, and a patient dose verification prediction result is generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data processing, and particularly relates to a VMAT plan patient dose verification prediction method and system based on Bayesian model uncertainty quantification. BACKGROUND

[0002] To ensure the accuracy of radiotherapy plan dose, traditional measurement method is time-consuming and laborious. Existing researches use plan complexity feature parameters or beam fluence map to develop AI-driven PSQA prediction models to improve efficiency, but the prediction accuracy and uncertainty warning need to be considered for promotion. The uncertainty quantification of AI model in current radiotherapy has not been paid enough attention. When the model confidence is low, it may produce unreliable prediction, which endangers the safety of patients. Therefore, it is necessary to quantify the uncertainty and introduce human intervention.

[0003] AI model uncertainty includes model (cognitive) uncertainty and data (random) uncertainty. The former is caused by insufficient model knowledge learning, and the data can reduce it. The latter is caused by inherent noise of data and cannot be avoided. The theoretical basis of uncertainty quantification is the Bayesian method. Early Gaussian process is limited in application due to high calculation cost. MCDropout generates uncertainty estimates in the validation phase, which has small overhead and is effective. Although self-aggregation method can improve the stability of the model, it is less applied due to the large data set of deep learning and the challenge of computing resources.

[0004] Uncertainty quantification is increasingly applied in target segmentation and other aspects of radiotherapy, but it is less applied in dose prediction and PSQA. In related research, Nguyen et al. used deep ensemble learning and MCDropout to quantify the uncertainty of dose distribution prediction. Yang et al. introduced MCDropout into the fluence map model to realize the uncertainty prediction of PSQA, but there are some shortcomings such as not fully reflecting the complex parameters of VMAT and not dealing with the "pass / fail" classification problem.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background disclosed by the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide a VMAT plan patient dose verification prediction method and system based on Bayesian model uncertainty quantification, at least to overcome the problems existing in the prior art to some extent, by collecting VMAT data containing TPS flux map and 54 field characteristics (including new MLC parameters), dividing the data set and labeling. Construct a multi-modal classification model, fuse the feature output classification result, and train Adam. Combined with the uncertainty of the Monte Carlo algorithm, determine T=10. After LASSO dimensionality reduction, build MLP to predict GPR with weighted MAE. Define the clinical threshold classification test set, and intervene when the threshold is exceeded. Combine EPID evaluation to generate results.

[0007] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the application.

[0008] According to one aspect of the present application, a VMAT plan patient dose verification prediction method based on Bayesian model uncertainty quantification is provided, comprising: collecting VMAT plan field data, including TPS calculated flux map and field complexity feature parameters, dividing training set, validation set and test set, determining PSQA classification label and GPR regression label based on γ analysis; based on the field complexity features and the field flux map, a multi-modal classification model is constructed, combined with the EPID measurement results, the classification performance is evaluated through the ROC curve, the confusion matrix, the sensitivity and the specificity, and the trained multi-modal classification model is generated, wherein the one-dimensional feature branch extracts features through the Conv1D layer with different convolution kernels, the two-dimensional image branch extracts features through the Conv2D layer, and the classification result is output after fusion through three fully connected layers, and the model is trained using the Adam optimizer; using the trained multi-modal classification model, combined with the Monte Carlo method, the data uncertainty, the model uncertainty and the total uncertainty are calculated through Monte Carlo sampling, the stability under different sampling times is evaluated, the clinically applicable sampling times are determined, and the preset clinical uncertainty threshold is generated based on the MMF-CA curve and the CC / IU curve analysis; based on the output of the last fully connected layer of the multi-modal classification model, the LASSO regression dimensionality reduction is performed, the classification probability and the corresponding uncertainty are used as inputs, the MLP model is constructed for predicting the GPR value, the self-defined weighted MAE loss function is used to train the model, the GPR prediction performance is evaluated through MAE and RMSE, and the trained GPR prediction model is generated; based on the preset clinical uncertainty threshold, the test set samples are classified, the samples exceeding the threshold trigger manual intervention, the GPR value of the samples is estimated using the trained GPR prediction model, and the comprehensive patient dose verification prediction result is generated.

[0009] In another aspect of the present application, a VMAT planned patient dose verification prediction device based on Bayesian model uncertainty quantification is characterized in that it comprises: an acquisition module for collecting VMAT planned field data, including TPS calculated flux maps and field complexity feature parameters, dividing training set, verification set and test set, determining PSQA classification labels and GPR regression labels based on gamma analysis; a processing module for constructing a multi-modal classification model based on field complexity features and field flux maps, combining EPID measurement results, evaluating classification performance through ROC curve, confusion matrix, sensitivity and specificity, generating a trained multi-modal classification model, wherein a one-dimensional feature branch extracts features through a Conv1D layer with different convolution kernels, a two-dimensional image branch extracts features through a Conv2D layer, and the fused features are output through three fully connected layers to generate a classification result, and the model is trained using an Adam optimizer; using the trained multi-modal classification model, combining a Monte Carlo method, calculating data uncertainty, model uncertainty and total uncertainty through Monte Carlo sampling, evaluating stability under different sampling frequencies, determining clinically applicable sampling frequencies, and generating a preset clinical uncertainty threshold based on MMF-CA curve and CC / IU curve analysis; based on the output of the last fully connected layer of the multi-modal classification model, reducing dimensions through LASSO regression, combining classification probability and corresponding uncertainty as input, constructing an MLP model for predicting GPR value, training the model using a custom weighted MAE loss function, evaluating GPR prediction performance through MAE and RMSE, and generating a trained GPR prediction model; based on the preset clinical uncertainty threshold, classifying test set samples, triggering manual intervention for samples exceeding the threshold, estimating the GPR value of the samples using the trained GPR prediction model, and generating a comprehensive patient dose verification prediction result.

[0010] According to another aspect of the present application, an electronic device is characterized in that it comprises: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to implement the above-mentioned VMAT planned patient dose verification prediction method based on Bayesian model uncertainty quantification.

[0011] The application provides a VMAT plan patient dose verification prediction method and system based on Bayesian model uncertainty quantification. The method comprises the following steps: collecting VMAT data containing TPS flux map and 54 field characteristics (containing newly added MLC speed and acceleration parameters), dividing data sets, and determining labels based on gamma analysis. A multi-modal classification model is constructed, a one-dimensional branch uses different Conv1D to extract features, a two-dimensional branch extracts flux map features through Conv2D, and a classification result is output after fusion. The model is trained by using an Adam optimizer. Combined with the variation Monte Carlo, the sampling number T is set to 3, the uncertainty is calculated, and T is determined to be 10 which is suitable for the actual clinical environment. After LASSO dimension reduction, the classification probability and the uncertainty are combined to construct an MLP, and a self-defined weighted MAE is used to train and predict a GPR. The clinical threshold is determined based on the curve, the test set is classified, manual intervention is performed when the threshold is exceeded, and the performance is evaluated in combination with EPID measurement to generate the result.

[0012] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flowchart of a VMAT plan patient dose verification prediction method based on Bayesian model uncertainty quantification is shown.

[0014] Figure 2 An analysis result schematic diagram of the number of Monte Carlo sampling times T in uncertainty quantification is shown.

[0015] Figure 3 A structure schematic diagram of a VMAT plan patient dose verification prediction device based on Bayesian model uncertainty quantification is shown. DETAILED DESCRIPTION

[0016] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0017] The VMAT plan patient dose verification prediction method based on Bayesian model uncertainty quantification according to the exemplary embodiments of the present application is described below in conjunction with Figure 1 It should be noted that the following application scenarios are only used to facilitate the understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect. On the contrary, the embodiments of the present application are applicable to any applicable scenario.

[0018] In one embodiment, the application further provides a VMAT plan patient dose verification prediction method and system based on Bayesian model uncertainty quantification. Figure 1A flowchart of a VMAT plan patient dose verification prediction method based on Bayesian model uncertainty quantification is shown schematically according to an embodiment of the application.

[0019] S101, collect the field data of the VMAT plan.

[0020] In an embodiment, the field data of 477 full-arc VMAT plans are collected, which are all from the plans designed by Eclipse TPS and irradiated by Trilogy linear accelerator. In an actual application scenario, when a hospital department of radiotherapy carries out related research, it extracts the VMAT plan field data in a specific time period from its treatment system, which covers treatment plans of various tumor types, providing a rich data basis for subsequent analysis.

[0021] The TPS calculated fluence map and field complexity feature parameters of each field are collected as inputs of the prediction model. The field complexity features newly added on the basis of the original research include the proportion of MLC leaf speed in the 2-3 cm / s interval (S 2-3 ), the proportion of acceleration in the 6-8 cm / s 2 interval (A 6-8 ), and the proportion of acceleration in the 8-20 cm / s 2 interval (A 8-20 ), a total of 54 complexity feature parameters are included. For example, when analyzing a specific VMAT plan, the fluence map of the plan is exported through the Eclipse TPS system, and the S 2-3 is calculated to be 0.3 (indicating that 30% of the MLC leaf speed is in the 2-3 cm / s interval), A 6-8 is 0.15, A 8-20 is 0.08, and the like.

[0022] The 477 field data collected are divided into training set and validation set according to the ratio of 7:3, and additionally 48 fields are collected as a prospective model test set. Taking a certain research as an example, 334 fields are randomly selected from the 477 fields as the training set for the initial training of the model; the remaining 143 fields are used as the validation set for adjusting and optimizing the hyperparameters of the model and evaluating the performance of the model under different parameter settings; and the additional 48 fields are used as the test set for the final evaluation of the generalization ability of the model on unseen data.

[0023] Measurement-based PSQA (Patient Dosage Validation) was performed using an Electronic Field Imaging Device (EPID), with gamma analysis completed in the PortalDosimetry module of EclipseTPS. Three gamma standards—3% / 3mm, 3% / 2mm, and 2% / 2mm—were selected for patient dose validation. Dosage differences were globally normalized using the maximum dose point, and the minimum dose threshold was set to 10%. Intervention limits were set at 95%, 95%, and 90%, respectively. Exceeding the intervention limit was considered "pass," and falling below it was considered "fail," thus determining the PSQA classification label. For the GPR regression label, the specific gamma pass rate was directly used as the label. For example, if a field had a gamma pass rate of 98% calculated using the 3% / 3mm gamma standard, its PSQA classification label was "pass," and its GPR regression label was 98%. If another field had a gamma pass rate of 88%, its PSQA classification label was "fail," and its GPR regression label was 88%.

[0024] S102, based on the complexity features of the shooting field and the flux map of the shooting field, constructs a multimodal classification model. Combined with the EPD measurement results, the classification performance is evaluated through ROC curve, confusion matrix, sensitivity and specificity to generate a trained multimodal classification model.

[0025] In one implementation, the input data for constructing the multimodal classification model includes one-dimensional field complexity features and a preprocessed two-dimensional field flux map. The one-dimensional field complexity features include 54 parameters, and based on previous research, the proportion of MLC blade velocities of 2-3 cm / s and accelerations of 6-8 cm / s² are added. 2 and 8-20cm / s 2 The preprocessed two-dimensional radiation field flux map was converted from DICOM format to a NumPy array, normalized to the [0,1] interval by Min-Max, zero-padding to 1199×1024, center-cropped to 1040×520, and downsampled to 64×32. One-dimensional radiation field complexity features (54 parameters): In VMAT (Volume Modulated Radiotherapy) planning, radiation field complexity directly affects dose delivery accuracy. These 54 features characterize the radiation field from dimensions such as "dose distribution control" and "mechanical motion characteristics," specifically including the basic parameters from previous studies (51 parameters): covering MU value (monitor units, jump count, reflecting the total irradiation dose), FieldX / Y (the size of the radiation field on the X / Y axis, determining the irradiation range), MLC blade positioning accuracy, and radiation field arc length. These parameters are core control quantities in TPS (Treatment Planning System) design planning. For example, in VMAT planning, FieldX / Y needs to adapt to the pelvic anatomy to ensure tumor coverage while protecting the intestines.

[0026] New motion features (3): S 2-3(MLC blade velocity ratio 2-3 cm / s), MLC (Multi-Leaf Collimator) is a key component of dynamic shaping rays, and its velocity directly affects the smoothness of the blades' "dose distribution tracking motion." In the plan, if S... 2-3 =0.3, indicating that the blades moved at a relatively stable speed (2-3 cm / s) during 30% of the movement, ensuring dosage accuracy while avoiding mechanical vibration. A 6-8 (acceleration 6-8cm / s²) 2 (proportion), A 8-20 (acceleration 8-20cm / s²) 2 (Proportion), acceleration reflects the severity of blade start-up, shutdown, and speed changes. A 6-8 =0.15 indicates that the blades undergo "medium-speed variation" during 15% of their motion, while A 8-20 =0.08 represents "rapid start-stop" in 8% of the motion. High acceleration may cause blade positioning errors, which need to be addressed through feature quantization to help the model identify "motion risks".

[0027] The flux map is a "ray intensity distribution map" calculated by TPS, directly reflecting the dose delivery pattern. The core of preprocessing is enabling the computer to efficiently "understand" this "dose map." DICOM to NumPy array conversion: DICOM is a standard format for medical images, storing metadata such as pixel values, patient information, and spatial location data from the flux map. Converting it to a NumPy array (the core structure of Python numerical computation) allows deep learning frameworks (such as TensorFlow and PyTorch) to process it. For example, a flux map DICOM file exported from Eclipse TPS, after being read by the pydicom library, can be converted into a two-dimensional array of shape [height, width], where each element is a ray intensity value.

[0028] Min-Max Normalization Among them, I raw I represents the original pixel value. max and I min These represent the maximum and minimum values ​​of a single flux map. Flux maps have a large range of pixel values ​​(e.g., 0-2000), and directly inputting them into the model can lead to training instability due to numerical differences. Normalizing to the [0,1] interval ensures that flux maps from different plans have a "uniform numerical scale," making it easier for the model to learn features. For example, the original value I of a pixel... raw ==1200, if the flux graph global I max =1500 (high dose area), I min =300 (low-dose region), normalized I norm =0.75 indicates that the dose intensity at this location is "relatively high", and the probability of it corresponding to the tumor target area is higher.

[0029] The VMAT project's flux map size is inconsistent (affected by field size and TPS settings), requiring a fixed input dimension to ensure the model can stably process data. If the original image is 1000×900, zero values ​​are symmetrically padded along the edges (similar to adding black borders to small photos) to ensure all samples are the same size, facilitating batch processing by the model. After padding, the image edges are "dose-free regions" (zero values), which do not affect the feature extraction of the central effective dose distribution.

[0030] The "effective dose distribution" of the flux map is concentrated in the central region (the irradiation range of the radiation field), and the zero values ​​filled at the edges are meaningless. Cropping the central 1040×520 pixels focuses key information and reduces model computation. For example, in the planned flux map, the dose distribution of the tumor target area and the intestinal protection area are both in the central region; cropping preserves the core features. Further compressing the image size balances "feature preservation" and "computational efficiency." A small size of 64×32 allows the model to complete training within a reasonable time while preserving the "macroscopic pattern" of the dose distribution (such as the shape of the high-dose area and the range of the low-dose area). Downsampling can be achieved through pooling operations (such as max pooling), merging information from adjacent pixels and preserving key trends.

[0031] In summary, one-dimensional feature quantification quantifies the "control parameters of field design and motion", while two-dimensional flux maps depict the "spatial pattern of dose delivery". The preprocessed data allows the multimodal model to simultaneously learn "how the field is designed" and "how the dose is actually distributed", ultimately accurately determining whether the PSQA is passed.

[0032] The multimodal classification model structure is designed as follows: The one-dimensional feature branch uses Conv1D layers with kernel sizes of 2 and 4 to extract local and global features respectively. Batch normalization is applied after each convolutional layer, and the convolutional features are concatenated and then processed by max pooling and Dropout. The two-dimensional image branch uses Conv2D layers with 3×3 kernels to extract spatial features. After convolution, batch normalization and ReLU activation are applied, followed by max pooling and flattening. The one-dimensional feature branch targets the complex features of the 54-dimensional field of view (including newly added MLC motion parameters). Multi-scale correlation features are extracted using Conv1D layers with different kernel sizes. The specific processing logic is as follows: Conv1D layers and multi-scale feature extraction: Conv1D layers with kernel sizes of 2 and 4 are used; Conv1D layer with kernel size 2: Focusing on local correlations, such as adjacent features (e.g., MLC velocity S). 2-3 With acceleration A 6-8 The coordinated changes of blade speed can be used to capture the dependencies between short-range parameters (e.g., whether the increase in blade speed is accompanied by a surge in acceleration).

[0033] Conv1D layer with convolution kernel 4: captures global correlations and covers features spanning 4 (such as MU value, FieldX, S). 2-3 A 8-20This refers to the long-range collaborative pattern of "total dose - field size - mechanical motion" in the reverse mapping field design. In a certain VMAT program, the layer with convolution kernel 2 can identify "S" in its 54-dimensional features. 2-3 With A 6-8 "positive correlation" (acceleration fluctuates significantly as speed increases); layers with convolution kernel 4 can capture "MU value = 500, FieldX = 15cm, S 2-3 =0.3, A 8-20 =0.08” combination mode (high dose, wide radiation field, more stable leaf movement).

[0034] Each Conv1D layer is followed by batch normalization: by standardizing the feature distribution (mean 0, variance 1), the problem of "large differences in numerical scales of different features" is solved (e.g., MU values ​​are usually in the hundreds, while S values ​​are much larger). 2-3 (0-1) to accelerate model convergence. Convolutional features are concatenated and then max-pooled (pooling size 2): the concatenated multi-scale features are downsampled to retain key features (such as peaks in local associations and salient patterns in global associations) and reduce data dimensionality. Dropout (0.4): 40% of neurons are randomly dropped during training to avoid the model from over-relying on a certain feature combination (such as "only through S"). 2-3 (Judging the complexity of the shooting field) enhances the generalization ability.

[0035] The 2D image branch extracts spatial features from the preprocessed 64×32 flux map (reflecting the spatial distribution of radiation intensity) through Conv2D layers and subsequent operations, specifically as follows: The Conv2D layer with 3×3 convolutional kernels focuses on local spatial correlations, such as the "dose change trend of a certain pixel and its eight surrounding pixels," capturing key features in the flux map such as high-dose region boundaries and dose gradients (e.g., the "high-dose core" of the planned tumor target area and the "low-dose attenuation zone" of the surrounding normal tissue). In the 64×32 flux map, the 3×3 convolutional kernel can identify the "continuity between the pixel value 0.8 (x=20, y=10) and the pixel value 0.75 to its right (x=20, y=11)," corresponding to the smooth dose transition at the edge of the target area; it can also capture the "abrupt change between the pixel value 0.9 (x=30, y=15) and the pixel value 0.3 below it (x=31, y=15)," corresponding to the dose drop at the field boundary.

[0036] Activation functions and pooling processes are detailed below: Batch normalization and ReLU activation: Batch normalization stabilizes the distribution of features after convolution; the ReLU activation function sets negative values ​​to 0, enhancing nonlinear expressions (e.g., highlighting features in high-dose regions and suppressing meaningless low-dose noise). Max pooling (2×2): Downsampling is performed on the convolutional feature map, retaining the maximum value within the 2×2 region (e.g., the peak value in high-dose regions), compressing data dimensionality while enhancing key spatial features. Flattening operation: The pooled two-dimensional feature map is converted into a one-dimensional vector, preparing for subsequent fusion with one-dimensional feature branches. For example, a 64×32 flux map, after convolution and pooling, yields a 32×16 feature map, which, after flattening, forms a 512-dimensional one-dimensional vector.

[0037] The one-dimensional branch and the two-dimensional branch extract features from the two dimensions of "correlation of field design parameters" and "spatial pattern of dose distribution" respectively. Multi-scale convolution and regularization are used to ensure the representativeness and stability of the features, providing a complementary information basis for the subsequent fusion layer and improving the accuracy of PSQA classification.

[0038] The outputs of the one-dimensional feature branch and the two-dimensional image branch are concatenated into a joint feature, which is then processed through three fully connected layers. Finally, the output layer, containing two units and a Softmax activation function, outputs the binary classification result. The one-dimensional feature branch targets the 54-dimensional field complexity features (including the newly added MLC motion parameters) and extracts multi-scale correlation features through Conv1D layers with different convolutional kernels. The specific processing logic is as follows: Conv1D layers with convolutional kernel sizes of 2 and 4 are used. The Conv1D layer with convolutional kernel size 2 focuses on local correlations, while the Conv1D layer with convolutional kernel size 4 captures global correlations.

[0039] Each Conv1D layer is followed by batch normalization, which addresses the problem of "large differences in numerical scales of different features" by standardizing the feature distribution (mean of 0, variance of 1). (For example, MU values ​​are usually in the hundreds, while S values ​​are much larger.) 2-3 (0-1) to accelerate model convergence. Convolutional features are concatenated and then max-pooled (pooling size 2): the concatenated multi-scale features are downsampled to retain key features (such as peaks in local associations and salient patterns in global associations) and reduce data dimensionality. Dropout (0.4): 40% of neurons are randomly dropped during training to avoid the model from over-relying on a certain feature combination (such as "only through S"). 2-3 (Judging the complexity of the shooting field) enhances the generalization ability.

[0040] The 2D image branch extracts spatial features from the preprocessed 64×32 flux map (reflecting the spatial distribution of ray intensity) through the Conv2D layer and subsequent operations. Specifically, the 3×3 convolutional kernel Conv2D layer focuses on local spatial correlations, such as the "dose change trend of a certain pixel and its 8 surrounding pixels", and captures key features such as the boundary of the high-dose region and the dose gradient in the flux map (e.g., the "high-dose core" of the planned tumor target area and the "low-dose attenuation zone" of the surrounding normal tissue).

[0041] Batch normalization and ReLU activation: Batch normalization stabilizes the distribution of features after convolution; the ReLU activation function sets negative values ​​to 0, enhancing nonlinear expressions (e.g., highlighting features in high-dose regions and suppressing meaningless low-dose noise). Max pooling (2×2): Downsampling is performed on the convolutional feature map, retaining the maximum value within the 2×2 region (e.g., the peak value in high-dose regions), compressing data dimensionality while enhancing key spatial features. Flattening operation: The pooled two-dimensional feature map is converted into a one-dimensional vector, preparing for subsequent fusion with one-dimensional feature branches. For example, a 64×32 flux map, after convolution and pooling, yields a 32×16 feature map, which, after flattening, forms a 512-dimensional one-dimensional vector.

[0042] The one-dimensional branch and the two-dimensional branch extract features from the two dimensions of "correlation of field design parameters" and "spatial pattern of dose distribution" respectively. Multi-scale convolution and regularization are used to ensure the representativeness and stability of the features, providing a complementary information basis for the subsequent fusion layer and improving the accuracy of PSQA classification.

[0043] The Adam optimizer is used to train a model to classify the results of the VMAT project PSQA, where the result classification includes pass or fail. In multimodal classification model training, the Adam optimizer addresses the problems of slow convergence and susceptibility to local optima in complex feature learning by adaptively adjusting the learning rate (combining the advantages of momentum and adaptive gradient squaredness). The learning rate is set to 0.01, 0.005, or 0.001, and needs to be determined through hyperparameter optimization. For classification tasks with different γ standards (3% / 3mm, 3% / 2mm, 2% / 2mm), the learning rate that maximizes the validation set AUC (area under the ROC curve) and minimizes the risk of overfitting is selected. The effects of the learning rate are as follows: 0.01: accelerates model parameter updates in the initial stage, quickly approaching the optimal solution; 0.005: balances convergence speed and stability, suitable for feature learning of most complex field data; 0.001: finely adjusts parameters in the later stages, avoiding convergence oscillations caused by gradient fluctuations.

[0044] The core objective of model training is to achieve binary classification ("pass" or "fail") of the VMAT program's PSQA results. The judgment criteria strictly follow the gamma analysis standard: based on three gamma standards (3% / 3mm, 3% / 2mm, 2% / 2mm), when the gamma pass rate (GPR) exceeds the intervention limit (95%, 95%, 90%, respectively), the label is "pass" (1); when it is below the limit, it is "fail" (0). Under the 3% / 2mm standard, the gamma analysis of one VMAT program yielded a GPR of 96% (intervention limit 95%), and the model output "pass"; another program had a GPR of 88% (below 95%), and the output was "fail".

[0045] When using the Adam optimizer with a learning rate of 0.005, the model progressively learns the association between "shooting field complexity features + flux map" and PSQA results by minimizing the cross-entropy loss (which measures the difference between predicted probabilities and true labels). For example, for features including "high MLC acceleration ratio (A... 8-20 For samples with "0.2) + disorder in the high gradient region of the flux map", the model will strengthen the prediction weight of the "failed" label.

[0046] After training, the model's performance on the validation set was evaluated using sensitivity, specificity, and AUC. 3% / 3mm standard: AUC = 0.8526, sensitivity 83.33% (proportion of correctly identified "failed" samples), specificity 90.38% (proportion of correctly identified "passed" samples); 3% / 2mm standard: AUC = 0.8926, sensitivity 93.33%, specificity 76.84%; 2% / 2mm standard: AUC = 0.8704, sensitivity 94.74%, specificity 73.63%. These results demonstrate that the model trained with the Adam optimizer can effectively distinguish between PSQA "passed" and "failed" samples, maintaining high sensitivity even under stringent gamma standards (such as 2% / 2mm), meeting the clinical need for a "no missed high-risk plan."

[0047] The classification performance of the multimodal classification model was evaluated using ROC curves, confusion matrices, sensitivity, and specificity, based on actual GPR values ​​measured by the Electronic Field Imaging Device (EPID). The classification performance of the multimodal classification model was evaluated using the following methods, based on actual GPR values ​​measured by the EPD:

[0048] ROC curve: Receiver Operating Characteristic Curve, with the False Positive Rate (FPR) on the x-axis and the True Positive Rate (TPR) on the y-axis. Where FP is the number of false positive samples (the number of samples that the model predicted as "failed" but actually "passed"), and TN is the number of true negative samples (the number of samples that the model predicted as "passed" but actually "passed"). Where TP is the number of true positive samples (samples that the model predicted as "failed" but actually failed), and FN is the number of false negative samples (samples that the model predicted as "passed" but actually failed). In the validation set under the 3% / 3mm standard, assuming the model prediction results are compared with the actual EPD measurement results, we get TP=5, FN=1, FP=10, and TN=94. Therefore, FPR=9.62% and TPR=83.33%. By calculating FPR and TPR at different thresholds, ROC curves are plotted; the closer the curve is to the upper left corner, the better the model's classification performance.

[0049] A confusion matrix is ​​a matrix that compares the predictions of a classification model with the actual results, visually representing the number of TP, FN, FP, and TN. Using the previous example, the confusion matrix is... Sensitivity: Also known as the true positive rate (TPR), it reflects the model's ability to correctly identify "pass" samples. In the example above, the sensitivity is 83.33%. Specificity: This reflects the model's ability to correctly identify "failed" samples. In this example,

[0050] S103 utilizes a trained multimodal classification model, combined with the variational Monte Carlo method, to calculate data uncertainty, model uncertainty, and total uncertainty through Monte Carlo sampling, evaluate stability under different sampling numbers, determine the clinically applicable sampling number, and generate a preset clinical uncertainty threshold based on MMF-CA curve and CC / IU curve analysis.

[0051] In one implementation, a trained multimodal classification model is used in conjunction with the variational Monte Carlo method, setting the number of Monte Carlo sampling times T to 3, 5, 10, 15, 20, and 25. In the uncertainty quantification framework based on Bayesian models, when performing uncertainty analysis on a trained multimodal classification model using the variational Monte Carlo method, the setting of the number of Monte Carlo sampling times T directly affects the stability and computational efficiency of uncertainty estimation. This application explicitly sets the value of T to 3, 5, 10, 15, 20, and 25, with the core purpose of finding the optimal value balancing stability and efficiency by comparing the results under different sampling times.

[0052] From a technical perspective, the variational Monte Carlo method quantifies prediction uncertainty by simulating the probability distribution of model parameters through multiple samplings. Too few samplings (e.g., T=3) result in insufficient sample representativeness, potentially leading to large fluctuations in uncertainty estimates; while too many samplings (e.g., T=25) improve stability but significantly increase computation time (computation time increases roughly linearly with T). Therefore, setting multiple T values ​​allows for a systematic evaluation of the trade-off between these two factors.

[0053] Taking a VMAT project under the 3% / 2mmγ standard as an example, the project's field data contains 54-dimensional complex features (such as MLC blade velocity S). 2-3 =0.28, acceleration A 6-8 =0.12) and the preprocessed 64×32 flux plot. Uncertainty analysis was performed using three sampling times: T=3, T=10, and T=25. When T=3, the data uncertainty U was calculated using only 3 Monte Carlo samplings. d Model uncertainty U m The total uncertainty U, due to the small number of samplings, results in significant differences in uncertainty for each calculation (e.g., the total uncertainty may fluctuate between 0.06 and 0.09), indicating insufficient stability. When T = 10, the range of uncertainty fluctuations obtained through 10 samplings is reduced (the total uncertainty stabilizes between 0.07 and 0.08), and the calculation time is controlled within a clinically acceptable range (approximately 7 seconds per sample). When T = 25, the stability of the results is further improved (the total uncertainty fluctuation is less than 0.005), but the calculation time increases to approximately 18 seconds per sample, far exceeding the clinical efficiency requirements.

[0054] This application utilizes a systematic evaluation of the validation set (such as...). Figure 2 As shown in the figure, T=10 was finally determined as the optimal number of samplings because, under the γ standard of 3% / 2mm, it can ensure that the fluctuation of the total uncertainty (standard deviation) is reduced to below 0.005, and the average inference time is controlled at a reasonable level, thus achieving a balance between stability and computational efficiency.

[0055] Through multiple Monte Carlo samplings, based on the Bernoulli distribution characteristics of binary classification problems, and using the formula... Uncertainty in computational data. d This refers to data uncertainty, reflecting the dispersion of classification probabilities caused by "data fluctuations" within the sample itself. T refers to the number of Monte Carlo samplings (e.g., T=10). The classification probability of the t-th sample (model-predicted sample x) * Belongs to the real label y * The probability, (These are the model parameters sampled for the t-th time).

[0056] For a certain VMAT project firing field data (x * Set T = 10 Monte Carlo samplings, and assume the classification probability of each sampling is... As follows: When the number of samplings t is 1, The value is 0.92 × 0.08; when the number of sampling times t is 2, The value is 0.91 × 0.09; when the number of sampling times t is 3, The value is 0.93 × 0.07; when the number of sampling times t is 4, The value is 0.90 × 0.10; when the number of sampling times t is 5, The value is 0.94 × 0.06; when the number of sampling times t is 6, The value is 0.92 × 0.08; when the number of sampling times t is 7, The value is 0.93 × 0.07; when the number of sampling times t is 8, The value is 0.91 × 0.09; when the number of sampling times t is 9, The value is 0.92 × 0.08; when the number of sampling times t is 10, It is 0.93 × 0.07. Substituting into the formula, we can calculate that...

[0057] Through formula Computational model uncertainty; U m This refers to model uncertainty, which reflects the classification probability bias caused by "fluctuations in model parameters". The average classification probability of the model (the mean of the probabilities from T samplings, i.e.) First calculate the average classification probability Substitute into the formula to calculate

[0058] Through formula The total uncertainty, U, is calculated by taking into account the overall impact of "data fluctuations" and "model parameter fluctuations," and then substituting this into the U obtained from the above calculation. d ≈0.0726 and U m ≈0.000139. (Because U m <<U d The total uncertainty is approximately equal to U. d If the model parameters fluctuate greatly, U m It will significantly affect the results.

[0059] The volatility of total uncertainty and average inference time were evaluated under different sampling numbers. The results showed that as T increases, the volatility of total uncertainty gradually decreases, the computation time increases roughly linearly, and the balance between stability and computational efficiency is optimal when T=10.

[0060] S104 is based on the output of the last fully connected layer of the multimodal classification model. After dimensionality reduction by LASSO regression, and combined with the classification probability and corresponding uncertainty as input, an MLP model is constructed to predict GPR values. The model is trained using a custom weighted MAE loss function, and the GPR prediction performance is evaluated by MAE and RMSE to generate a trained GPR prediction model.

[0061] In one implementation, the output features of the last fully connected layer of the multimodal classification model are obtained, the learnable parameters of the model are frozen, and LASSO regression is used to filter the output features to achieve parameter dimensionality reduction. The output features are obtained from the last fully connected layer of the trained multimodal classification model (which integrates one-dimensional field complexity features and two-dimensional flux map features). These features are high-dimensional representations of the model's comprehensive learning of the input data, containing rich information related to the PSQA results. For example, in the VMAT project, after training, the last fully connected layer of the multimodal classification model outputs a 128-dimensional feature vector, which may encode a combination of field parameters (such as MLC blade velocity and acceleration-related features) and flux map spatial dose distribution features.

[0062] The learnable parameters of the multimodal classification model are frozen so that they are not updated in subsequent processes. This is because the model has already been trained on the PSQA classification task, and fixing the parameters ensures the stability of the extracted features, avoiding feature drift caused by parameter changes in the new task (GPR value prediction). Taking the previous VMAT project as an example, when obtaining the 128-dimensional feature vector, the model parameters are not adjusted to ensure the consistency of feature extraction. LASSO (Least Absolute Shrinkage and Selection Operator) regression is used to filter and reduce the dimensionality of the extracted high-dimensional features. LASSO regression introduces an absolute value penalty term into the regression model, making the coefficients of some features become 0, thereby achieving the purpose of feature selection and dimensionality reduction. For the 128-dimensional feature vector of the above VMAT project, after LASSO regression, at least three key features are selected. These features have a stronger correlation with the GPR value of the field of view, removing redundant information and reducing the computational complexity of subsequent processes.

[0063] The dimensionality-reduced features are combined with the classification probabilities and uncertainty results as input features, which are then standardized. The LASSO-reduced features output from the last fully connected layer of the multimodal classification model are then combined with the classification probabilities. Data uncertainty (U d Model uncertainty (U) mThe total uncertainty (U) is combined and normalized to the [0,1] interval using Min-Max. Specifically, the dimensionality-reduced features are [0.32, 0.15, 0.08] (MLC velocity ratio, acceleration ratio, etc.), and the classification probability and uncertainty are... U = 0.0726. The combined features are [0.32, 0.15, 0.08, 0.921, 0.0726]. After standardization, the feature values ​​remain in the [0, 1] range (no adjustment required).

[0064] An MLP model is constructed, comprising an input layer, three hidden layers, and an output layer. A single neuron, in conjunction with a sigmoid activation function, constrains the predicted value to the [0,1] interval. The MLP model (Multilayer Perceptron) used in this application to predict GPR values ​​takes the intermediate features, classification probabilities, and uncertainty results of a multimodal classification model as input. Through nonlinear transformation of the three hidden layers, it ultimately outputs a GPR prediction value constrained to the [0,1] interval (corresponding to a γ pass rate of 0%-100%).

[0065] Specifically, the structure of the input layer is as follows: Number of neurons: consistent with the dimensionality of the standardized input features (5 neurons in this example). Input feature composition: Features from the last fully connected layer of the multimodal classification model after dimensionality reduction via LASSO regression (such as key features fused with field complexity and flux map); Classification probability (average classification probability obtained through variational Monte Carlo sampling). Uncertainty results (data uncertainty U) d Model uncertainty U m The total uncertainty (U) is a factor. Preprocessing requires that all input features be standardized (e.g., Min-Max normalization to the [0,1] interval) to unify the units and improve the stability of model training.

[0066] The input features of a certain VMAT project are [0.32 (dimensionality reduction feature 1), 0.15 (dimensionality reduction feature 2), 0.08 (dimensionality reduction feature 3), 0.921]. [0.0726(U)] corresponds to 5 neurons in the input layer. The structure of the hidden layers is as follows: Hidden layer 1: 128 neurons, using the ReLU activation function, to capture high-order nonlinear relationships between input features (such as the synergistic effect of field complexity features and uncertainty), achieving preliminary feature extraction. Hidden layer 2: 64 neurons, using the ReLU activation function, further compressing and abstracting the output of hidden layer 1, focusing on core patterns strongly correlated with GPR values ​​(such as the association between high uncertainty samples and low GPR values). Hidden layer 3: 32 neurons, using the ReLU activation function, strengthening the representation ability of key features with fewer neurons, providing refined input for regression prediction in the output layer.

[0067] The ReLU activation function can effectively solve the gradient vanishing problem and is suitable for processing high-dimensional medical data; the design of decreasing depth and number of neurons in the three hidden layers balances feature learning ability and computational efficiency, which meets the requirement of this application to "achieve high-precision prediction within a clinically acceptable time".

[0068] The structure of the output layer is as follows: Number of neurons: 1 neuron. With the Sigmoid activation function, the output of hidden layer 3 is mapped to the [0,1] interval, directly corresponding to the GPR value (e.g., an output of 0.93 indicates a predicted γ pass rate of 93%). The output characteristics of the Sigmoid function are strictly matched with the physical range of γ pass rate (0%-100%), ensuring the clinical interpretability of the prediction results.

[0069] Using a custom weighted MAE loss function The model was trained to predict the GPR value of the shooting field. Prediction analysis was performed on both the full validation set and failed samples to evaluate model performance. Here, L is the weighted average absolute error loss, used to measure the deviation between the model's predicted value and the true value; a smaller value indicates a more accurate prediction. N is the total number of samples used to calculate the loss. i These are the base weights, used to weight samples whose GPR is less than a threshold. For classification probability weights and For uncertain weights, y i Let be the true field GRP value of the i-th sample. The model predicts the field GPR value for the i-th sample, using classification probability and uncertainty to guide the GPR prediction. i All are 5; for 3% / 3mm, and The ratios are 4:1 for 3% / 2mm and 2% / 2mm, and The ratios are 1:1.

[0070] base weight w i The determination is based on whether the sample GPR is less than a set threshold. If it is less than the threshold (i.e., the sample failed), w i =5; if greater than or equal to the threshold (through samples), w i =1 (default no additional weighting, can be adjusted according to actual needs). For example, for the 3% / 3mm standard, the GPR qualification threshold is assumed to be 95%. If a sample's GPR is 92% (less than 95%), then w i =5.

[0071] Classification probability weights and uncertainty weights To establish the proportional relationship based on the γ standard type. Taking the 3% / 3mm standard as an example, set up but (a is a positive real number); for 3% / 2mm and 2% / 2mm standards, Right now The true value y i To obtain true GPR data of the radiation field through actual dose verification measurement methods (such as EPD imaging), serving as a reference benchmark for model prediction. Predicted values. The output is obtained by forward propagation of the samples using the constructed MLP model (which includes an input layer, three hidden layers, and an output layer, with the output layer activated by Sigmoid). The output range is [0,1], and it is converted into a percentage form as needed (e.g., multiplied by 100 to get the corresponding GPR percentage).

[0072] The predictive performance of the GPR prediction model is evaluated using mean absolute error (MAE) and root mean square error (RMSE). The mean absolute error (MAE) is calculated using the formula... Where n is the number of samples in the test set, y i It is the actual GPR value of the i-th sample (obtained by EPID measurement). This is the predicted GPR value for the i-th sample. In the validation set, there are n = 143 samples. Assume the actual GPR value of the first sample is y1 = 0.90, and the predicted value is... The second sample has y² = 0.95. The average deviation between predicted and true values ​​is measured by averaging the absolute values ​​of error across all samples.

[0073] The formula for root mean square error (RMSE) is: Similarly, for the above n=143 samples, first calculate the square of the error for each sample, then calculate the average, and finally take the square root. For example, the square of the error for the first sample is (0.90-0.88). 2 =0.0004, second sample (0.95-0.93) 2 =0.0004, etc., RMSE penalizes samples with larger errors more severely and can more sensitively reflect the dispersion of predicted values.

[0074] S105 classifies the test set samples based on a preset clinical uncertainty threshold. Samples exceeding the threshold trigger manual intervention. The trained GPR prediction model is used to estimate the GPR value of the samples and generate comprehensive patient dose validation prediction results.

[0075] In one implementation, the clinical uncertainty threshold is determined based on the validation set using MMF-CA curves and CC / IU curves. The clinical uncertainty thresholds are 0.057 for 3% / 3mm, 0.014 for 3% / 2mm, and 0.080 for 2% / 2mm. Under the 3% / 3mm standard, the field data in the validation set are analyzed. By plotting the MMF-CA curve (with the manually measured MMF ratio on the horizontal axis and the clinical accuracy CA on the vertical axis) and the CC / IU curve (CC representing the probability that the model is confident in the correct classification, and IU representing the probability that the model is not confident in the outcome when the prediction is incorrect), the curve trends show that when the uncertainty threshold is between 0.057 and 0.240, the CC and IU curves exhibit a rapid increasing trend. This range is a key selection range for the clinically preset uncertainty threshold. Considering the clinical requirements for sensitivity (ensuring 100% sensitivity to protect patient safety), 0.057 is ultimately determined as the clinical uncertainty threshold under this standard.

[0076] The same curve analysis was performed on the validation set radiation field data under the 3% / 2mmγ standard. The results showed that when the uncertainty threshold was between 0.014 and 0.207, the CC and IU curves exhibited a rapid increasing trend. Based on the principle of ensuring clinical safety and ensuring that all model misclassification samples can be corrected, the minimum value within this range, 0.014, was selected as the clinical uncertainty threshold under this standard.

[0077] For the 2% / 2mmγ standard, analysis of the MMF-CA curves and CC / IU curves of the validation set radiation field data revealed a rapid increasing trend in both CC and IU curves when the uncertainty threshold was between 0.080 and 0.244. To ensure 100% clinical sensitivity, the minimum value within this range, 0.080, was selected as the clinical uncertainty threshold for this standard. These thresholds are crucial for model performance evaluation and subsequent decision-making. When the total uncertainty of a sample exceeds the corresponding threshold, manual intervention is triggered to ensure the accuracy of clinical classification.

[0078] A pre-trained multimodal classification model is used to classify the test set samples. This model integrates multimodal information such as field complexity features (e.g., MLC blade velocity and acceleration-related features) and flux map features, outputting a classification result ("pass" or "fail") for each test set sample. Simultaneously, the total uncertainty for each sample, previously calculated using the variational Monte Carlo method, is also considered.

[0079] The calculated total uncertainty is compared with a preset clinical uncertainty threshold. If the total uncertainty of a sample exceeds the threshold, it indicates that the model's classification result for that sample is of low reliability, requiring manual intervention. For example, under the 3% / 2mm standard, a sample in a test set is judged as "pass" by the multimodal classification model, but its calculated total uncertainty value is 0.02 (exceeding the threshold of 0.014 under this standard). In this case, it is necessary to manually re-examine the parameters, flux map, and other information of the radiation field, and it may even be necessary to re-measure and verify.

[0080] If the total uncertainty of a sample does not exceed the threshold, the model prediction result is directly adopted. For example, under the 3% / 3mm standard, for another VMAT planned firing field, the total uncertainty is 0.04 (less than the threshold of 0.057), and the model prediction is "pass", then the classification result is accepted.

[0081] The trained GPR prediction model is used to estimate the GPR values ​​of the test set samples to obtain predicted values. For a VMAT plan test set sample, the feature vector of the input GPR prediction model consists of three parts: first, the output features of the last fully connected layer of the multimodal classification model after dimensionality reduction by LASSO regression (the parameters of this classification model have been frozen to ensure feature stability); second, the classification probability of the sample obtained by the variational Monte Carlo method; and third, the data uncertainty, model uncertainty, and total uncertainty results calculated by the variational Monte Carlo method. These features are all standardized before input to unify the dimensions and improve the model's processing efficiency.

[0082] The GPR prediction model is a multilayer perceptron (MLP) structure, consisting of an input layer, three hidden layers, and an output layer. The number of neurons in the input layer matches the dimension of the standardized feature vector. The three hidden layers have 128, 64, and 32 neurons respectively, all employing the ReLU activation function to capture non-linear features in the data and achieve high-level feature extraction. The output layer has a single neuron, which, along with the Sigmoid activation function, constrains the predicted value to the [0,1] interval to match the distribution characteristics of GPR values.

[0083] When the feature vector of a VMAT test set sample is input into the model and processed by the aforementioned network structure, the output is the GPR prediction value for that sample. If the output value is 0.93, it indicates that the model estimates the gamma pass rate (GPR) of that field under the corresponding gamma standard to be 93%. This prediction result, combined with the model's uncertainty analysis, can provide a quantitative reference for clinical judgment on whether the treatment plan has passed dose validation.

[0084] By combining classification performance and GPR prediction performance, patient dose validation prediction results are generated. If the multimodal classification model exhibits high sensitivity (e.g., close to 1) and specificity (e.g., close to 1) under different gamma standards, and the GPR prediction model has low MAE and RMSE (e.g., MAE < 0.03, RMSE < 0.05), it indicates good overall model performance. Results are presented in report form, including performance metrics for each model under different gamma standards and the classification result ("pass" or "fail") for each patient's test set sample. For example, for a patient's VMAT plan, the report shows that the multimodal classification model judged it as "pass" under the 3% / 2mm standard.

[0085] This application collects VMAT project firing field data, including TPS computational throughput maps and 54 firing field complexity features (with newly added MLC velocity and acceleration-related parameters). Training, validation, and test sets are divided, and PSQA classification and GPR regression labels are determined based on gamma analysis. A multimodal classification model is constructed. The one-dimensional branch uses Conv1D with different convolutional kernels to extract local and global features, while the two-dimensional branch extracts throughput map spatial features through Conv2D. After fusion, the classification results are output through a fully connected layer and trained using the Adam optimizer.

[0086] Combining the variational Monte Carlo method, sampling times of T=3, 5, and 10 were set to calculate data, model, and total uncertainty, determining that T=10 is the most clinically applicable. Based on the classification model output, dimensionality reduction was performed using LASSO, and an MLP predictive GPR was constructed by combining classification probability and uncertainty, trained using a custom weighted MAE loss function. Clinical thresholds (e.g., 0.057 for 3% / 3mm) were determined based on MMF-CA and CC / IU curves. The test set was classified, and exceeding the threshold triggered manual intervention, generating comprehensive validation results.

[0087] In one implementation, such as Figure 3 As shown, this application also provides a VMAT plan patient dose validation and prediction device based on Bayesian model uncertainty quantification, comprising:

[0088] The acquisition module 301 is used to collect the shooting field data of the VMAT plan, including the TPS calculation throughput map and shooting field complexity feature parameters, divide the training set, validation set and test set, and determine the PSQA classification label and GPR regression label based on gamma analysis;

[0089] Processing module 302 is used to construct a multimodal classification model based on the complexity features and flux map of the shooting field. The one-dimensional feature branch extracts features through Conv1D layers with different convolutional kernels, while the two-dimensional image branch extracts features through Conv2D layers. After fusion, the models are output as classification results via three fully connected layers. The model is trained using the Adam optimizer, and its classification performance is evaluated using ROC curves, confusion matrices, sensitivity, and specificity, combined with EPID measurement results. Using the trained multimodal classification model, and in conjunction with the variational Monte Carlo method, data uncertainty, model uncertainty, and total uncertainty are calculated through Monte Carlo sampling to assess stability under different sampling numbers and determine clinical applicability. The sampling number is determined; based on the output of the last fully connected layer of the multimodal classification model, LASSO regression dimensionality reduction is performed, and the classification probability and corresponding uncertainty are used as inputs to construct an MLP model for predicting GPR values. A custom weighted MAE loss function is used to train the model, and the GPR prediction performance is evaluated by MAE and RMSE; based on a preset clinical uncertainty threshold, the test set samples are classified, and samples exceeding the threshold trigger manual intervention. The GPR prediction model is used to estimate the GPR values ​​of the samples, generating comprehensive patient dose verification prediction results. The preset clinical uncertainty threshold is obtained based on MMF-CA curve and CC / IU curve analysis.

[0090] The computer-readable storage medium provided in the above embodiments of this application and the VMAT plan patient dose verification and prediction method based on Bayesian model uncertainty quantification provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0091] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the VMAT plan patient dose verification and prediction method, electronic device, electronic device, and readable storage medium based on Bayesian model uncertainty quantification are basically similar to the embodiments of the VMAT plan patient dose verification and prediction method described above, and are therefore described simply. Relevant parts can be referred to in the descriptions of the embodiments of the VMAT plan patient dose verification and prediction method based on Bayesian model uncertainty quantification.

Claims

1. A method for patient dose validation and prediction in VMAT planning based on Bayesian model uncertainty quantification, characterized in that, include: Collect the firing field data of the VMAT project, including the TPS computation throughput map and firing field complexity feature parameters, divide the training set, validation set and test set, and determine the PSQA classification label and GPR regression label based on gamma analysis; Based on the complexity features and flux map of the shooting field, a multimodal classification model is constructed. Combined with the EPD measurement results, the classification performance is evaluated by ROC curve, confusion matrix, sensitivity, and specificity to generate a trained multimodal classification model. The one-dimensional feature branch extracts features through Conv1D layers with different convolution kernels, and the two-dimensional image branch extracts features through Conv2D layers. After fusion, the classification results are output through three fully connected layers. The Adam optimizer is used to train the model. Using a trained multimodal classification model and the variational Monte Carlo method, data uncertainty, model uncertainty and total uncertainty are calculated through Monte Carlo sampling. The stability under different sampling times is evaluated, the clinically applicable sampling times are determined, and a preset clinical uncertainty threshold is generated based on MMF-CA curve and CC / IU curve analysis. Based on the output of the last fully connected layer of the multimodal classification model, LASSO regression dimensionality reduction is performed, and the classification probability and corresponding uncertainty are combined as input to construct an MLP model for predicting GPR values. A custom weighted MAE loss function is used to train the model, and the GPR prediction performance is evaluated by MAE and RMSE to generate a trained GPR prediction model. Based on a preset clinical uncertainty threshold, the test set samples are classified. Samples exceeding the threshold trigger manual intervention. The trained GPR prediction model is used to estimate the GPR value of the samples and generate comprehensive patient dose validation prediction results.

2. The method as described in claim 1, characterized in that, Based on the complexity features and flux map of the shooting field, a multimodal classification model is constructed. Combining EPID measurement results, the classification performance is evaluated using ROC curves, confusion matrices, sensitivity, and specificity. A pre-trained multimodal classification model is then generated, including: The input data for constructing the multimodal classification model includes one-dimensional field complexity features and a preprocessed two-dimensional field flux map. The one-dimensional field complexity features include 54 parameters, including the proportion of newly added MLC blade velocities of 2-3 cm / s and accelerations of 6-8 cm / s². 2 and 8-20cm / s 2 The preprocessed two-dimensional field flux map is converted from DICOM format to NumPy array, normalized to the [0,1] interval by Min-Max, zero-padded to 1199×1024, centered clipped to 1040×520, and downsampled to 64×32. The multimodal classification model structure is designed as follows: the one-dimensional feature branch uses Conv1D layers with kernel sizes of 2 and 4 to extract local and global features respectively. Batch normalization is applied after each convolutional layer, and the convolutional features are concatenated and then processed by max pooling and Dropout. The two-dimensional image branch uses Conv2D layers with 3×3 kernels to extract spatial features. After convolution, batch normalization and ReLU activation function are applied, followed by max pooling and flattening operations. The outputs of the one-dimensional feature branch and the two-dimensional image branch are concatenated into joint features, which are then processed through three fully connected layers. Finally, the binary classification result is output through an output layer containing two units and a Softmax activation function. The Adam optimizer is used to train the model to classify the VMAT program PSQA results, where the result classification includes pass or fail; By combining the EPID measurement results, the classification performance is evaluated through ROC curves, confusion matrix, sensitivity, and specificity, and a well-trained multimodal classification model is generated.

3. The method as described in claim 1, characterized in that, Using a trained multimodal classification model and the variational Monte Carlo method, data uncertainty, model uncertainty, and total uncertainty are calculated through Monte Carlo sampling. Stability is evaluated under different sampling numbers to determine the clinically applicable sampling number. Pre-defined clinical uncertainty thresholds are generated based on MMF-CA curves and CC / IU curve analysis, including: Using a trained multimodal classification model and the variational Monte Carlo method, the number of Monte Carlo sampling times T is set to 3, 5, 10, 15, 20, and 25. Through multiple Monte Carlo samplings, based on the Bernoulli distribution characteristics of binary classification problems, and using the formula... Calculate the uncertainty of the data, U d To represent data uncertainty, T is the number of Monte Carlo samplings, t is the index of a single Monte Carlo sampling, and y * For real labels, x * For the input sample, Let be the model parameters for the t-th sampling. Parameters at the tth sampling time Below, the model applies the input sample x * Predicted as true label y * The probability of; Through formula Calculate the uncertainty of the model, U m Due to model uncertainty, The model is based on the input x * Predicted as true label y * The average probability; Through formula Calculate the total uncertainty; The volatility of total uncertainty and average inference time were evaluated under different sampling numbers. The results showed that as T increases, the volatility of total uncertainty gradually decreases, the computation time increases roughly linearly, and the balance between stability and computational efficiency is optimal when T=10. Preset clinical uncertainty thresholds were generated through MMF-CA curve and CC / IU curve analysis. The clinical uncertainty thresholds were 0.057 for 3% / 3mm, 0.014 for 3% / 2mm, and 0.080 for 2% / 2mm.

4. The method as described in claim 1, characterized in that, Based on the output of the last fully connected layer of the multimodal classification model, LASSO regression dimensionality reduction is applied. Combining the classification probability and corresponding uncertainty as input, an MLP model is constructed to predict GPR values. A custom weighted MAE loss function is used to train the model, and the GPR prediction performance is evaluated using MAE and RMSE. The resulting trained GPR prediction model includes: Obtain the output features of the last fully connected layer of the multimodal classification model, freeze the learnable parameters of the model, and use LASSO regression to filter the output features to achieve parameter dimensionality reduction; The dimensionality-reduced features are combined with the classification probability and uncertainty results as input features, and the input features are then standardized. Construct an MLP model, which includes an input layer, three hidden layers, and an output layer. In this model, a single neuron, in conjunction with the Sigmoid activation function, constrains the predicted value to the interval [0,1]. Using a custom weighted MAE loss function The model was trained to predict the GPR value of the shooting field. Prediction analysis was performed on both the full validation set and failed samples to evaluate the model performance. Here, L is a custom weighted average absolute error loss function, N is the total number of samples in the training set, and y... i Let be the true GPR value of the i-th sample. w is the GPR prediction value of the model for the i-th sample. i These are the base weights, used to weight samples whose GPR is less than a threshold. and These are the contributions of increasing classification probability and uncertainty to the loss, respectively, and how classification probability and uncertainty guide GPR prediction. i All are 5; for 3% / 3mm, and The ratios are 4:1 for 3% / 2mm and 2% / 2mm, and The ratios are 1:1 respectively; The GPR prediction performance is evaluated using MAE and RMSE, and a trained GPR prediction model is generated.

5. The method as described in claim 1, characterized in that, Based on a preset clinical uncertainty threshold, the test set samples are classified. Samples exceeding the threshold trigger manual intervention. A trained GPR prediction model is used to estimate the GPR value of the samples, generating comprehensive patient dose validation prediction results, including: The test set samples are classified using a trained multimodal classification model. The total uncertainty is compared with a preset threshold. Samples exceeding the threshold trigger manual intervention, while samples below the threshold directly use the model prediction results. The trained GPR prediction model is used to estimate the GPR values ​​of the test set samples to obtain the predicted values. By combining the classification results, the results after clinical uncertainty intervention, and the GPR prediction performance, a comprehensive patient dose validation prediction result is generated.

6. A device for verifying and predicting patient dose in VMAT planning based on Bayesian model uncertainty quantification, characterized in that, The device includes: The acquisition module is used to collect the firing field data of the VMAT plan, including the TPS computation throughput map and firing field complexity feature parameters, divide the training set, validation set and test set, and determine the PSQA classification label and GPR regression label based on gamma analysis; The processing module constructs a multimodal classification model based on the complexity features and flux map of the shooting field. Combining EPD measurement results, it evaluates classification performance using ROC curves, confusion matrices, sensitivity, and specificity, generating a trained multimodal classification model. The one-dimensional feature branch extracts features through Conv1D layers with different convolutional kernels, while the two-dimensional image branch extracts features through Conv2D layers. The fused features are then processed through three fully connected layers to output the classification result. The model is trained using the Adam optimizer. Using the trained multimodal classification model, and in conjunction with the variational Monte Carlo method, data uncertainty, model uncertainty, and total uncertainty are calculated through Monte Carlo sampling. The stability under different sampling numbers is evaluated to determine clinical applicability. The sampling frequency and analysis based on MMF-CA curves and CC / IU curves generate a preset clinical uncertainty threshold. Based on the output of the last fully connected layer of the multimodal classification model, LASSO regression dimensionality reduction is performed. Combining classification probability and corresponding uncertainty as input, an MLP model is constructed to predict GPR values. A custom weighted MAE loss function is used to train the model, and the GPR prediction performance is evaluated using MAE and RMSE to generate a trained GPR prediction model. Based on the preset clinical uncertainty threshold, the test set samples are classified. Samples exceeding the threshold trigger manual intervention. The trained GPR prediction model is used to estimate the GPR values ​​of the samples, generating a comprehensive patient dose validation prediction result.

7. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the VMAT plan patient dose verification prediction method based on Bayesian model uncertainty quantification as described in any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the VMAT plan patient dose verification and prediction method based on Bayesian model uncertainty quantification as described in any one of claims 1 to 5.

Images