Multi-granularity uncertainty-based plan quality assurance prediction method

Abstract

The invention discloses a plan quality assurance prediction method based on multi-granularity uncertainty, and the method comprises the following steps: S1, collecting data, and carrying out the preprocessing; and S2, processing the data based on a multi-granularity regression prior network model to obtain Gamma passing rate GPR prediction and dose difference prediction DDP. According to the method, the dose difference and the distribution of the Gamma passing rate can be deduced, modeling is carried out under a specific statistical hypothesis, and the credibility of deep learning application in plan quality assurance can be improved by using the framework and the network architecture.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a planning quality assurance prediction method based on multi-granularity uncertainty. Background Art

[0002] Deep learning automated plan quality assurance (PSQA) aims to reduce clinical resource requirements. Predicting the dose difference metric (gamma pass rate) and its distribution is critical to ensuring the safety and effectiveness of radiotherapy. However, current studies ignore the uncertainty limits in model predictions, limiting their credibility in real clinical settings. Summary of the invention

[0003] In view of the above technical features, the present invention provides a planning quality assurance prediction method based on multi-granularity uncertainty, comprising the following steps: S1: data collection and preprocessing; S2: The data is processed based on the multi-granularity regression prior network model to obtain the gamma pass rate GPR prediction and dose difference prediction DDP.

[0004] Furthermore, step S1 includes the following sub-steps: S11: Acquire 2D dose image data; S12: resampling the 2D dose image data to obtain an image with consistent spatial resolution; S13: Cutting, removing redundant background; S14: Perform normalization processing to ensure the consistency of input multi-granularity regression prior network model data.

[0005] Further, step S2 includes establishing a basic Bayesian model, and the basic Bayesian model includes the following calculation method: ; in, Indicates that given input The probability of regression variable y under the condition of and training data D, represents the likelihood of y given the model parameters µ; Indicates that given input The posterior distribution of the model parameter µ given the hyperparameter ϕ; The posterior distribution of the hyperparameter ϕ given the training data D.

[0006] Furthermore, the basic Bayesian model uses the Normal-Wishart distribution as a priori, and the calculation method is: ; ; in, is the distribution of the target variable y given input x and parameter θ; µ is the mean, Λ is the precision matrix; {µ,Λ} is obtained from the function f(x; θ).

[0007] Furthermore, step S2 also includes generating a preset distribution and bringing it into the Bayesian model to obtain a parameterized form of the output distribution: ; ; ; in, represents the Normal-Wishart distribution, which is a composite distribution consisting of a conditional normal distribution on mean µ and a Wishart distribution on the precision matrix Λ; represents a parameterization of the Normal-Wishart distribution over mean µ and precision matrix Λ; the distribution parameters {m, L, κ, ν} are obtained from the function f(x;θ).

[0008] Furthermore, step S2 also includes calculating a parameterized expression of uncertainty: ; ; ; in, represents the model uncertainty, which is calculated based on the determinant of the precision matrix Λ; represents the entropy of the predicted probability distribution p(y|µ,Λ) over µ and Λ, indicating the total uncertainty; Indicates data uncertainty.

[0009] Furthermore, step S2 also includes outputting probability distribution parameters according to the parameterized form of the output distribution and calculating the training loss function form: .

[0010] Furthermore, step S2 also includes outputting relevant parameters through the deep learning network, thereby completing the GPR prediction and DDP and their prediction confidence measurement, and achieving consistency of the two task predictions: ; ; ; ; in, is the maximum likelihood of the GPR distribution.

[0011] The beneficial effects of the present invention are that the present invention can infer the distribution of dose differences and gamma pass rates and perform modeling under specific statistical assumptions. By using this framework and network architecture, the credibility of deep learning applications in plan quality assurance can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0013] Figure 1 It is a flow chart of the present invention; Figure 2 The network architecture of the present invention; Figure 3 Build a flowchart for your model. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0015] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0016] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0017] See also Figure 1 , a planning quality assurance prediction method based on multi-granularity uncertainty (MGUQ) includes the following steps: S1: data collection and preprocessing; S2: The data is processed based on the multi-granularity regression prior network model to obtain the gamma pass rate GPR prediction and dose difference prediction DDP.

[0018] In this embodiment, step S1 includes the following sub-steps: S11: Acquire 2D dose image data; S12: resampling the 2D dose image data to obtain an image with consistent spatial resolution; S13: Cutting, removing redundant background; S14: Perform normalization processing to ensure the consistency of input multi-granularity regression prior network model data.

[0019] The dose image can be the calculated dose map of the treatment planning system (TPS), exported in DICOM format, and used as the input of the deep learning network. The original energy density map has different spatial resolutions and sizes, so several preprocessing steps are required. First, the 2D energy density map is resampled to a consistent spatial resolution of 1mm × 1mm, and then cropped to 512 × 512 pixels to remove redundant background. Finally, before entering the network, the pixel values ​​of the input image are rescaled to [0, 1] by Min-Max normalization to ensure the consistency of the input data.

[0020] In this embodiment, the multi-granularity regression prior network model is optimized for multi-task learning, allowing simultaneous training of DDP and GPR prediction tasks. Integrating granular uncertainty information can enhance the adaptability of the network to various PSQA scenarios. The equations mentioned below are common settings for prior network related work, which are regarded as prior assumptions; in this case, these equations are followed to lay the foundation for the proposed multi-granularity regression prior network:

[0021] ; in, Indicates that given input The probability of the regression variable y given the training data D, which represents the prediction uncertainty; represents the likelihood of y given the model parameters µ, which quantifies the uncertainty associated with the model parameters; Indicates that given input The posterior distribution of the model parameter µ under the condition of the input and hyperparameter ϕ, which represents the uncertainty of the model parameter under the condition of the input and hyperparameter; The posterior distribution of the hyperparameter ϕ given the training data D.

[0022] The basic assumption involves modeling the lower-granularity task DDP as a Gaussian distribution. Given that the GPR aggregates DDP points above a threshold, the probability distribution of the higher-granularity GPR tasks can be viewed as an average representation of the uncertainty in the DDP. Furthermore, the probability used for inference can be viewed as the probabilistic average of the likelihood probabilities. Therefore, it is reasonable to view the GPR distribution as a prior for the DDP. When dealing with Gaussian distributions with unknown mean and variance, the Normal-Wishart (NW) distribution is usually used as a prior:

[0023] ; ; where p(y|x, θ) is the distribution of the target variable y given input x and parameters θ. The distribution is assumed to be a multivariate normal distribution with mean µ and precision matrix Λ, where {µ,Λ} is obtained from the function f(x; θ).

[0024] Furthermore, the Normal-Wishart distribution is defined as: ; It is a composite distribution consisting of a conditional normal distribution over mean µ and a Wishart distribution over precision matrix Λ.

[0025] ; represents the parameterization of the Normal-Wishart distribution with mean µ and precision matrix Λ given input x and parameters θ, where the distribution parameters {m, L, κ, ν} are obtained from the function f(x; θ).

[0026] ; The predicted distribution p(y|x, θ) is calculated by taking the expected value of the parameter distribution p(µ,Λ|x, θ), and the resulting distribution is expressed as , and if v is infinite, the T distribution converges to the normal distribution.

[0027] ; represents model uncertainty, denoted by , which is calculated based on the determinant of the precision matrix Λ. Among them, represents the entropy of the predicted probability distribution p(y|µ,Λ) on µ and Λ, representing the total uncertainty, Indicates data uncertainty.

[0028] Model uncertainty, total uncertainty, and expected model uncertainty in the prior network can also be expressed in terms of variance: ; in, represents model uncertainty. Further, represents the overall uncertainty, and Demonstrate uncertainty in expected data.

[0029] , specifies the relationship for inducing uncertainty precision in the Normal-Wishart prior distribution, where is a normal distribution.

[0030] The predicted distribution of the high-granularity task GPR is the prior distribution of the low-granularity task DDP. Therefore, according to Bayesian theory, a variational distribution is constructed based on a deep neural network, denoted as p(µ,Λ|x,θ), to approximate the posterior distribution p(µ,Λ|Ω), and the approximate value is used to obtain the lower bound of evidence as the loss function for training DDP. The formula is as follows:

[0031] ; ; ; ; ; In the above formula, the complex relationship between granularity levels, both lower and higher, is illustrated. Here, L1 is the maximum likelihood of the GPR distribution, and then the GPR distribution is used as the prior distribution of DDP. By optimizing the posterior distribution of DDP, L2 and L3 are obtained. L2 can be regarded as the Lddp loss, which means that if only the DDP task is focused on, only this loss exists. L3 simulates the consistency of the two granularity tasks, hereinafter referred to as the consistency loss. By applying consistency loss on uncertainty types of different granularities, a more comprehensive and accurate uncertainty estimate can be achieved. This supervised information enables the neural network to learn more reasonable and informative representations within limited data.

[0032] In order to implement the above method, the present invention also provides a network architecture, such as Figure 2As shown, the first component is a set of encoders consisting of the GPR predictions and a Student-t distribution derived from the predicted NW distribution parameters. The Student-t distribution handles uncertainty and takes parameters such as location factor (LF), scale factor (SF), and degrees of freedom (DF). LF represents the mean in a Gaussian distribution and determines the central value or mean of the distribution. SF controls the variance of the Student-t distribution. DF affects the shape and tail of the Student-t distribution. The next component is the decoder, which generates the output for the DDP and processes the encoded input to generate the output prediction. Where (Mean) represents the mean of the distribution and Sigma represents the standard deviation of the output distribution, representing the variation in the model predictions. The architecture integrates different encoders to capture different uncertainties in PSQA data.

[0033] also, Figure 2 The use of Unet in combination with ResNet

[50] is also demonstrated. ResNet is a classic and efficient structure that has been widely used in many medical image tasks. Unet and ResNet are suggested to be used as the backbone as an example because this model architecture is flexible enough to use other models. When the neural network reaches a certain depth, increasing the number of layers does not bring any improvement because the gradient disappears and the network performance decreases. ResNet proposes the idea of ​​residual learning to alleviate the above problem by fitting the residual function F(x)=H(x)-x (where x and H(x) are the input and output of the residual block, respectively). ResNet18 is used as the feature extraction backbone of the network, which mainly consists of four stages, each of which includes two cascaded residual blocks. The residual block includes a convolutional layer, an instance normalization layer, and a ReLU activation layer for learning the residual mapping. The backbone can be adjusted to other pre-trained networks as needed. The network consists of two identical backbone networks. Features are extracted from each layer of the backbone network, and then the features of the two networks are fused through interactive links. The output of the last layer is then fed into the regression head separately to predict the prior distribution parameters of the mean (normal distribution) and variance (Wishart distribution). These parameters then synthesize the overall prior distribution, namely the Normal-Wishart distribution. The Student-t distribution of the GPR is then calculated based on this prior and the Gaussian likelihood. On the other hand, we feed the feature list into the U-Net decoder to output the prediction results of the DDP.

[0034] In general, if Figure 3As shown, the present invention begins with a modeling phase, in which multiple data sets are used, including training, validation, and calibration sets. The trained MGUQ model is then used to establish criteria for uncertainty thresholds or ranges. This phase requires evaluating and measuring the level of ambiguity associated with making clinical decisions. Once the level of uncertainty is determined, it is compared to a preset threshold. When the uncertainty is above the threshold, it indicates a high degree of uncertainty in the decision. In this case, manual intervention may be required to make a more informed and reliable choice. Conversely, if the level of uncertainty is below the specified threshold, the choice is deemed acceptable and no manual intervention is required. This indicates that the ambiguity of the decision-making process has been reduced and the predetermined clinical criteria have been met. In summary, the flowchart illustrates a systematic process for measuring uncertainty in clinical decision-making, which can improve understanding and control of uncertainty by implementing appropriate intervention thresholds.

[0035] The stability of MGUQ in PSQA was evaluated by multiple indicators, including clinical accuracy, manual measurement score ratio, correct certainty, mean absolute error, sensitivity, and specificity. The experiment showed that in terms of PSQA uncertainty measurement, MGUQ proved that the correlation of multi-granularity tasks can optimize the performance of uncertainty quantification; thereby enhancing safety and interpretability. This shows its potential for real-world clinical application, and this study has promoted the development of this field.

[0036] For the aforementioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily required by the present application.

[0037] The above embodiments describe the basic principles and main features of the present invention and the advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the changes and modifications made by those skilled in the art shall be within the scope of protection of the appended claims of the present invention without departing from the spirit and scope of the present invention.

Claims

1. A planning quality assurance prediction method based on multi-granularity uncertainty, characterized in that: The steps include: S1: data collection and preprocessing; S2: The data is processed based on the multi-granularity regression prior network model to obtain the gamma pass rate GPR prediction and dose difference prediction DDP.

2. A planning quality assurance prediction method based on multi-granularity uncertainty as claimed in claim 1, characterized in that: Step S1 includes the following sub-steps: S11: Acquire 2D dose image data; S12: resampling the 2D dose image data to obtain an image with consistent spatial resolution; S13: Cutting, removing redundant background; S14: Perform normalization processing to ensure the consistency of input multi-granularity regression prior network model data.

3. A planning quality assurance prediction method based on multi-granularity uncertainty as claimed in claim 2, characterized in that: Step S2 includes establishing a basic Bayesian model, and the basic Bayesian model includes the following calculation method: ； in, Indicates that given input The probability of regression variable y under the condition of and training data D, represents the likelihood of y given the model parameters µ; Indicates that given input The posterior distribution of the model parameter µ given the hyperparameter ϕ; The posterior distribution of the hyperparameter ϕ given the training data D.

4. A planning quality assurance prediction method based on multi-granularity uncertainty as claimed in claim 3, characterized in that: The basic Bayesian model uses the Normal-Wishart distribution as a priori, and the calculation method is: ； ； in, is the distribution of the target variable y given input x and parameter θ; µ is the mean, Λ is the precision matrix; {µ,Λ} is obtained from the function f(x; θ).

5. A planning quality assurance prediction method based on multi-granularity uncertainty as claimed in claim 4, characterized in that: Step S2 also includes generating a preset distribution and bringing it into the Bayesian model to obtain a parameterized form of the output distribution: ； ； ； in, represents the Normal-Wishart distribution, which is a composite distribution consisting of a conditional normal distribution on mean µ and a Wishart distribution on the precision matrix Λ; represents the parameterization of the Normal-Wishart distribution over mean µ and precision matrix Λ; the distribution parameters {m, L, κ, ν} are obtained from the function f(x;θ).

6. A planning quality assurance prediction method based on multi-granularity uncertainty as claimed in claim 5, characterized in that: Step S2 also includes calculating a parameterized expression for the uncertainty: ； ； ； in, represents the model uncertainty, which is calculated based on the determinant of the precision matrix Λ; represents the entropy of the predicted probability distribution p(y|µ,Λ) over µ and Λ, indicating the total uncertainty; Indicates data uncertainty.

7. A planning quality assurance prediction method based on multi-granularity uncertainty as claimed in claim 6, characterized in that: Step S2 also includes outputting probability distribution parameters according to the parameterized form of the output distribution and calculating the training loss function form: 。 8. A planning quality assurance prediction method based on multi-granularity uncertainty as claimed in claim 7, characterized in that: Step S2 also includes outputting relevant parameters through the deep learning network to complete the GPR prediction and DDP and their prediction confidence measurement, and achieve consistency of the two task predictions: ； ； ； ； in, is the maximum likelihood of the GPR distribution.

Images