Gamma passing rate prediction conformal risk control method

Abstract

The invention discloses a gamma passing rate prediction conformal risk control method, and the method comprises the following steps: S1, collecting data, and carrying out the preprocessing; s2, inputting the plan table modal features and the image modal features into a multi-modal deep network to obtain quantile features; s3, performing dose difference quantile decoding on the quantile characteristics to obtain a dose difference matrix quantile and a gamma passing rate quantile; and S4, conformal prediction is carried out, and conformal risk control of dose difference is carried out. According to the method, the prediction safety is relatively high, and the consistency of uncertainty estimation is relatively high.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a gamma pass rate prediction conformal risk control method. Background Art

[0002] Most existing gamma pass rate predictions only focus on global performance, but not enough on the safety of predictions. Machine learning methods based on statistical learning usually have large prediction variances on some data, but current machine learning-based gamma pass rate prediction algorithms do not measure this uncertainty.

[0003] Most existing uncertainty measurement algorithms usually only use symmetric probability modeling and ignore asymmetric risks. However, the use of probability alone in gamma pass rate prediction is insufficient, because predicting a qualified plan as unqualified only requires redesigning the plan, while predicting an unqualified plan as qualified carries the risk of medical accidents.

[0004] Most existing uncertainty algorithms based on deep learning usually suffer from the problem of overestimation of confidence and statistical inconsistency with actual test results when measuring uncertainty. Summary of the invention

[0005] In view of the above technical features, the present invention provides a gamma pass rate prediction conformal risk control method, comprising the following steps:

[0006] S1: data collection and preprocessing;

[0007] S2: Input the planning form modality and image modality features into the multimodal deep network to obtain quantile features;

[0008] S3: Decode the quantile features into dose difference quantiles to obtain the dose difference matrix quantiles and gamma pass rate quantiles;

[0009] S4: Conformal prediction and conformal risk control of dose differences.

[0010] Furthermore, step S1 includes the following sub-steps:

[0011] S11: Acquire 2D dose image data;

[0012] S12: resampling the 2D dose image data to obtain an image with consistent spatial resolution;

[0013] S13: Cutting, removing redundant background;

[0014] S14: performing normalization processing;

[0015] S15: Perform feature extraction and feature conversion to obtain basic features Zgpr.

[0016] Furthermore, step S2 includes the following sub-steps:

[0017] S21: Input the basic feature Zgpr into the GPR quantile decoder and use the regression model to predict the three quantiles Yddp;

[0018] S22: The basic feature Zddp is input to the dose difference quantile decoder, and the dose difference is predicted using the U-Net decoder, which converts the channel into one for regression through three convolutional layers;

[0019] S23: GPR quantile decoder outputs three quantiles of gamma pass rate qα low , q1 / 2 and qα up , ,The dose difference quantile decoder outputs the dose difference segmentation results and the corresponding uncertainty quantification.

[0020] Furthermore, step S3 includes the following sub-steps:

[0021] S31: Build a multi-granularity prior network architecture that allows simultaneous training of DDP and GPR prediction tasks;

[0022] S32: Quantile regression using Pinball loss, calculated as:

[0023]

[0024] Furthermore, step S4 includes the following sub-steps:

[0025] S41: Segmentation-based conformal prediction:

[0026] P(Y n+1 ∈C α (X n+1 ))≥1-α

[0027]

[0028] Furthermore, step S4 also includes the following sub-steps:

[0029] S42: Perform conformal risk control:

[0030]

[0031] Furthermore, step S4 also includes the following sub-steps:

[0032] S43: Perform conformal risk control of dose differences:

[0033]

[0034] in, α is the user-set risk, and λ is the calibration coefficient; it is calculated as follows:

[0035] First place formula:

[0036]

[0037] The second formula in order, the definition of the λ set:

[0038]

[0039] The third formula is the calibration coefficient λ, which is used last. After obtaining the calibration coefficient λ, the calibration coefficient λ should be substituted into T λ To perform a calibration:

[0040]

[0041] α w is the weighted quantile transformed into .

[0042] The beneficial effects of the present invention are that the prediction of the present invention is relatively safe and the consistency of uncertainty estimation is relatively high. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] 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.

[0044] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0045] 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.

[0046] 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.

[0047] 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.

[0048] See also Figure 1 , a gamma pass rate prediction conformal risk control method, comprising the following steps:

[0049] S1: data collection and preprocessing;

[0050] S2: Input the planning form modality and image modality features into the multimodal deep network to obtain quantile features;

[0051] S3: Decode the quantile features into dose difference quantiles to obtain the dose difference matrix quantiles and gamma pass rate quantiles;

[0052] S4: Conformal prediction and conformal risk control of dose differences.

[0053] In this embodiment, step S1 acquires a dose image, which exhibits different spatial resolutions and sizes, so several preprocessing steps are required. First, the 2D dose 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] through Min-Max normalization to ensure the consistency of the input data.

[0054] Each sample contains two modality data: image modality and tabular data. Specifically, the image modality is a 512×512 matrix representing the image modality. All element values ​​are between 0 and 1, indicating the relative dose intensity; the tabular data is a 33-dimensional vector, and all elements are complexity indicators representing the tabular modality. The specific target variable of the GPR prediction task is a three-dimensional vector corresponding to three common standards in clinical practice. GPR represents the gamma pass rate. The specific target variable of the DDP task is a 512×512 matrix, which is the same as the image modality input. DDP represents the dose difference. These definitions enable us to effectively model the PSQA (deep learning automated plan quality assurance) task and perform corresponding predictions and verifications.

[0055] Then, feature extraction is performed through the network flow: input data, image modality data (e.g., dose plan arrays) are processed through an image encoder (e.g., ResNet) to extract features. Tabular modality data (e.g., complexity indicators) are processed through a tabular encoder to extract features.

[0056] Feature extraction, the image encoder converts the image modality data into a feature vector of dimension Nc × W × H. The table encoder converts the table modality data into a feature vector of dimension Nt·W·H (where W and H are the last two dimensions of the image modality).

[0057] Feature transformation, the extracted image features and tabular features are concatenated from different paths to form Nc+Nt-dimensional Zgpr. The tabular modality features of shape Batch×(Nt·W·H) are reshaped into Batch×Nt×W×H and then concatenated with the image modality data. The quantile feature fuser merges these features into a unified base feature Zddp.

[0058] Step S2 is specifically as follows: Decoding is performed through the dose difference quantile decoder. To handle asymmetric risks, several quantile forms are set, including distribution and non-distribution quantiles. The basic feature Zgpr is input to the GPR quantile decoder, and the three quantiles Yddp are predicted using the regression model; the basic feature Zddp is input to the dose difference quantile decoder, and the dose difference is predicted using the U-Net decoder. Three convolutional layers convert the channel into one for regression; the GPR quantile decoder outputs three quantiles of the Gamma pass rate {qαlow, q1 / 2, qαup}. The dose difference quantile decoder outputs the dose difference segmentation result and the corresponding uncertainty quantification.

[0059] Step S3 is as follows: Calculation is performed through a multi-granularity regression prior network architecture, which is optimized for multi-task learning and allows simultaneous training of DDP and GPR prediction tasks. Quantile regression with Pinball loss is used. Since the data is extremely unbalanced in the task of radiotherapy plan verification, most samples pass, while only a small number of samples fail, and we have an asymmetric risk preference, most uncertainty methods are symmetric distribution modeling (some references), so we adopt quantile regression modeling without assuming distribution to evaluate uncertainty.

[0060] The purpose of quantile regression is to estimate a given quantile level τ of the target variable Y given the value x of the feature variable X. In other words, we are interested in building a model Approximate conditional quantile distribution function y = Q τ (Y|X). One strategy for estimating such a model is to minimize the Pinball loss:

[0061]

[0062] To demonstrate this, we write down the expected loss:

[0063]

[0064] Differentiate the above expression:

[0065]

[0066] Setting the above expression to zero can find the quantile The absolute loss corresponds to It is related to the estimated conditional median and corresponds to the MAE loss. The model applies the pinball loss to output the lower bound, median, and upper quantile To quantify asymmetric uncertainty, usually α up Greater than α low .

[0067] Step S4 specifically includes: performing conformal prediction:

[0068] Deep quantile regression alone usually fails to achieve strict uncertainty quantification. Simply optimizing a loss function that incorporates an uncertainty measure often leads to overestimation or underestimation problems. This approach can only serve as a heuristic measure of uncertainty, which severely limits the application of deep learning in clinical practice. On the other hand, due to the complexity of the situation, doctors sometimes need quantile values ​​that the model has not been trained on. Although some works have attempted to input quantiles into the model to obtain outputs of any quantile, many studies have shown that this approach cannot produce satisfactory quantile estimates. Unlike the deep uncertainty measurement algorithms used during training mentioned above, conformal learning uses post-hoc calibration, which can provide model-independent finite sample guarantees for the predictions of machine learning models and is considered to achieve strict uncertainty quantification.

[0069] Deep conformal learning is generally implemented based on split conformal prediction (split CP). Split conformal prediction requires a calibration dataset independent of the training, validation, and test datasets, as well as an acceptable significance level α∈(0,1) set by the user. Split conformal prediction (Split CP) first needs to define the inconsistency score S(x,y) (usually a measure of prediction error) and construct a prediction set Cα(Xn+1) for the new test sample Xn+1. Its mathematical form is:

[0070] P(Y n+1 ∈C α (X n+1 ))≥1-α

[0071]

[0072] The only assumptions are that the calibration and test datasets form an exchangeable set of random variables (a weaker condition than independent and identically distributed) and that both are independent of the training dataset.

[0073] Conformal Risk Control,While conformal prediction has achieved rigorous uncertainty metrics, it does not take into account the highly asymmetric risks in radiotherapy plan verification. Misclassifying a valid plan as invalid often requires only additional computing resources to redesign the plan, while misclassifying an invalid plan as valid may lead to potential medical accidents and threaten patient lives. In addition, the distribution of IMRT plan data is very unbalanced, with most plans being valid and only a few being invalid. In this case, deep learning models tend to predict effectiveness. Therefore, more sophisticated inconsistency metrics are needed to better meet clinical needs. Conformal Risk Control (CRC) generalizes the idea of ​​conformal prediction to this setting, ensuring that the prediction set keeps the expected risk below a user-predefined level α. The mathematical form is:

[0074]

[0075] Unlike the training loss function that compares the true value and the point prediction, CRC l considers the loss function that compares the true value and the set prediction. If l is defined as a non-coverage indicator function (binary loss), it is equal to 1 when the non-coverage condition is met and 0 otherwise:

[0076]

[0077] Then CRC recovers to the definition of CP, because the expected value of the indicator function of an event is equal to the probability of the event occurring. Therefore, CP can be regarded as a special case of CRC. The probability guarantee of CRC is based on the following minimal assumptions: the test dataset DE and calibration dataset DC samples are independent and identically distributed, and the training dataset DT is independent of the test dataset DE and calibration dataset DC.

[0078] CRC requires the user to set a loss function l to represent risk or error mathematically based on the application requirements, and then set a maximum tolerable risk α. The loss function l is not used as a training loss, which means that applying CRC does not directly affect the model training or prediction algorithm. Instead, the loss function l allows the user to "encode the concept of error" in their predictions. For clarity, the loss function l will be referred to as the conformal loss. Using the retained calibration data, it is possible to estimate Obtain finite sample, model-independent, and marginal guarantees for CRC.

[0079] Finally, performing conformal risk control for dose differences, for the DDP quantile model that generates pixel scores for the matrix regression predictor, for input X we have:

[0080] q(X)=(qk(X))ij, ij∈[1,..H]×[1,..W], k∈[alow, 1 / 2, αup]};

[0081] where ij corresponds to the location of the dose point in the planning matrix and k corresponds to the three quantiles. Given λ and quantiles We define the calibration interval C λ (X) and the calibration function T λ (q(X)) is as follows:

[0082]

[0083] It should be noted that in the DDP task, are matrices. In this context, we keep the above definitions and λ as scalars, where operations between matrices use matrix addition and subtraction, and operations between matrices and scalar λ use scalar multiplication on matrices. Finally, we use to represent the pixel interval in the corresponding index of the plane array.

[0084]

[0085] For the GPR task, we calibrated the three standards separately, so C λ All operations of (X) are scalar operations. The larger the λ, the larger the interval:

[0086] make Consider the loss function We assume that l takes values ​​in the bounded interval (-∞,B], where We define:

[0087]

[0088] We consider a series of IMRT plans and their corresponding predicted target values ​​(X i , Y i )i=1 n+1 The first n samples constitute our calibration set D C , the n+1th sample is regarded as a test sample. We denote L i (λ)=(Cλ(X i ), Y i ),

[0089] The calibration set is intended to estimate the correct value and ensure that the risk remains below the maximum tolerable risk level. Given the maximum tolerable risk level We define:

[0090]

[0091] Since l is usually a ratio or indicator function, the upper limit B of l is usually 1. In implementation, we rearrange the terms of the above equation to obtain the following form:

[0092]

[0093] For clarity, we define So the expression can be written as:

[0094]

[0095] Let L i (λ) is non-increasing, right continuous, and bounded by B<+∞. Suppose there exists λmax such that Li(λmax)≤α. Further suppose L1(λ),...,LN+1(λ) constitutes an exchangeable sequence, then:

[0096] Due to L i The monotonicity of (λ) with respect to λ, It can be found by performing an exponential search on the parameter λ, with an error range defined by the user as ∈; it is calculated as follows:

[0097] First place formula:

[0098]

[0099] The second formula in order, the definition of the λ set:

[0100]

[0101] The third formula is the calibration coefficient λ, which is used last. After obtaining the calibration coefficient λ, the calibration coefficient λ should be substituted into T λ To perform a calibration:

[0102]

[0103] α w is the weighted quantile transformed into .

[0104] 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.

[0105] 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 gamma pass rate prediction conformal risk control method, characterized in that: The steps include: S1: data collection and preprocessing; S2: Input the planning form modality and image modality features into the multimodal deep network to obtain quantile features; S3: Decode the quantile features into dose difference quantiles to obtain the dose difference matrix quantiles and gamma pass rate quantiles; S4: Conformal prediction and conformal risk control of dose differences.

2. A gamma pass rate prediction conformal risk control method 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: performing normalization processing; S15: Perform feature extraction and feature conversion to obtain basic features Zgpr.

3. A gamma pass rate prediction conformal risk control method as claimed in claim 2, characterized in that: Step S2 includes the following sub-steps: S21: Input the basic feature Zgpr into the GPR quantile decoder and use the regression model to predict the three quantiles Yddp; S22: The basic feature Zddp is input to the dose difference quantile decoder, and the dose difference is predicted using the U-Net decoder, which converts the channel into one for regression through three convolutional layers; S23: GPR quantile decoder outputs three quantiles of gamma pass rate and The dose difference quantile decoder outputs the dose difference segmentation results and the corresponding uncertainty quantification.

4. A gamma pass rate prediction conformal risk control method as claimed in claim 3, characterized in that: Step S3 includes the following sub-steps: S31: Build a multi-granularity prior network architecture that allows simultaneous training of DDP and GPR prediction tasks; S32: Quantile regression using Pinball loss, calculated as:

5. A gamma pass rate prediction conformal risk control method as claimed in claim 4, characterized in that: Step S4 includes the following sub-steps: S41: Segmentation-based conformal prediction: P(Y n+1 ∈C α (X n+1 ))≥1-a 6. A gamma pass rate prediction conformal risk control method as claimed in claim 5, characterized in that: Step S4 also includes the following sub-steps: S42: Perform conformal risk control:

7. A gamma pass rate prediction conformal risk control method as claimed in claim 6, characterized in that: Step S4 also includes the following sub-steps: S43: Perform conformal risk control of dose differences: in, α is the user-set risk, and λ is the calibration coefficient; it is calculated as follows: First place formula: The second formula in order, the definition of the λ set: The third formula is the last formula used for the calibration coefficient λ. After obtaining the calibration coefficient λ, substitute the calibration coefficient λ into T λ To perform a calibration: α w is the weighted quantile transformed into .

Images