Optimization method for boron neutron capture therapy dose calculation based on random forest model

Abstract

The invention discloses an optimization method for boron neutron capture therapy dose calculation based on a random forest model, and the technical scheme is as follows: firstly, setting a plurality of groups of irradiation variables for dose calculation to obtain a sample data set, and then extracting key information from sample data and simplifying the sample data; and training a random forest model by using the extracted key information and the simplified sample data. After training is completed, results of all possible irradiation variables are predicted by using a random forest model, and a change graph of dose effect indexes along with different irradiation variables is made. And then, verifying the reliability of the random forest model, and confirming that the error is within an allowable range. And finally, after the reliability of the random forest model is verified, the dose optimization result is verified to be actually improved. Prediction is carried out through dose calculation in combination with the random forest regression model, the optimal irradiation variable is efficiently and accurately found, and the method has the advantages of being short in consumed time, small in consumed computing power, simple in process and the like.

Description

Technical Field

[0001] The present invention relates to an optimization method for boron neutron capture therapy dose calculation, and particularly to an optimization method for boron neutron capture therapy dose calculation based on a random forest model. Background Art

[0002] BNCT mainly injects a selective boron-containing drug into the human body and enriches it in tumor cells. After being irradiated by thermal neutrons, 10 the high-energy, short-range α particles released by the 10B reaction can kill tumor cells and avoid damaging normal tissues. The treatment planning system (TPS) calculates the irradiation process of ray particles to obtain an optimal configuration in terms of neutron energy spectrum, patient positioning, and dose distribution in tumors and normal tissues. Its accurate and efficient calculation is crucial for the efficacy of BNCT.

[0003] Currently, the treatment plans on the market mainly consist of three modules: pre-processing, calculation, and post-processing. By reconstructing the three-dimensional geometric model, it is completely described as the irradiated tissue and the region of interest; calculating the particle transport process to give the dose values of the irradiated tissue and the region of interest; displaying and analyzing the calculation results, and then giving the treatment plan.

[0004] The above existing treatment plans involve a large amount of calculation in each link, especially in the calculation and post-processing links. The following problems generally exist when giving the optimal irradiation variables: long time consumption, high computing power consumption, complex process, and poor accuracy in the dose calculation process, which cannot meet the market demand for high efficiency and precision. Summary of the Invention

[0005] In view of the above-mentioned defects existing in the prior art, the present invention provides an optimization method for boron neutron capture therapy dose calculation based on a random forest model. This method is an accurate and efficient method for optimizing dose calculation. By combining dose calculation with a random forest regression model for prediction, it can efficiently and accurately find the optimal irradiation variables, with the advantages of short time consumption, low computing power consumption, and simple process.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] An optimization method for boron neutron capture therapy dose calculation based on a random forest model, comprising the following steps:

[0008] Step 1: Obtaining dose calculation and sample data sets: Set multiple groups of irradiation variables, each group of irradiation variables including boron drug concentration, irradiation azimuth, and source strength distribution. Through pre-processing, Monte Carlo calculation program calculation, and post-processing for conventional dose calculation, the corresponding dose calculation results are obtained respectively as the sample data sets;

[0009] Step 2: Train the random forest model: Obtain the doses corresponding to several groups of different irradiation variables as sample data. After extracting and simplifying the key information from the sample data, use the sample data with the key information extracted and simplified to train the random forest model;

[0010] Step 3: Use the random forest model for prediction: After training, use the random forest model to predict the results of all possible irradiation variables;

[0011] Step 4: Draw a graph of the dose effect index changing with different irradiation variables: From the results of all possible irradiation variables obtained by prediction, calculate the dose effect index according to formula (1) as the basis for measuring the dose effect, and thus draw a graph of the dose effect index changing with different irradiation variables;

[0012]

[0013] where I is the dose effect evaluation index; are the doses corresponding to the tumor region percentage volumes of 99%, 50%, and 1% in the DVH diagram respectively; are the doses corresponding to the sensitive region percentage volumes of 99%, 50%, and 1% in the DVH diagram respectively;

[0014] Step 5: Verify the reliability of the random forest model: Randomly select a part of the irradiation variables, calculate the error between the dose results obtained through pre-processing, Monte Carlo calculation program calculation, and post-processing and the dose results predicted by the random forest model, and confirm whether the error is within the allowable range. If it exceeds the range, repeat Steps 1 - 3, and after training a random forest model with better prediction effect, verify the reliability again;

[0015] Step 6: Verify the dose optimization result: After verifying the reliability of the random forest model in Step 5, use the graph of the dose effect changing with different irradiation variables in Step 4 to find the irradiation variables corresponding to the highest point of the dose effect evaluation index in this graph. This set of irradiation variables is the theoretical optimal irradiation variables. Calculate the error between the dose calculation results obtained by pre-processing, Monte Carlo calculation program calculation, and post-processing of this set of theoretical optimal irradiation variables and the prediction results of the random forest model to verify the accuracy of the prediction of this set of theoretical optimal irradiation variables. On the premise of accuracy, verify that this optimization result has a real improvement through the comparison of the effects before and after optimization.

[0016] Preferably, the random forest model uses a multi-output regression random forest model.

[0017] Preferably, the sample data is divided into an 80% model training set and a 20% model test set.

[0018] Preferably, the criterion for verifying the reliability of the random forest model is:

[0019]

[0020] Among them, x i (i = 1, 2, 3) are the doses corresponding to the tumor regional percentage volumes of 1%, 50%, and 99% in the DVH diagram obtained by predicting a set of irradiation variables respectively; a i (i = 1, 2, 3) are the doses corresponding to the tumor regional percentage volumes of 1%, 50%, and 99% in the DVH diagram calculated by a set of irradiation variables respectively; y j (j = 1, 2, 3) are the doses corresponding to the sensitive regional percentage volumes of 1%, 50%, and 99% in the DVH diagram obtained by predicting a set of irradiation variables respectively; b j (j = 1, 2, 3) are the doses corresponding to the sensitive regional percentage volumes of 1%, 50%, and 99% in the DVH diagram calculated by a set of irradiation variables respectively.

[0021] The beneficial effects of the present invention are as follows:

[0022] (1) The dose results of each tissue are obtained by combining the calculation with the random forest model (in the form of a DVH diagram), which is accurate and efficient: by combining the prediction results of the machine learning - random forest model, the number of groups of irradiation variables that need to be calculated is effectively reduced, saving time and reducing computing power consumption;

[0023] (2) The random forest model has a memory function, and the more training data, the higher the accuracy. Moreover, the prediction results have been verified for reliability and are relatively accurate; the theoretical optimal irradiation variables are obtained from the dose calculation effect change curve and verified by calculation, confirming that there is a real improvement in the treatment accuracy and treatment effect under the condition of the same number of calculation examples. Description of the Drawings

[0024] Figure 1 is the flow chart of the optimization method of the present invention;

[0025] Figure 2 is the effect diagram of training the random forest model;

[0026] Figure 3 is the diagram of the dose effect changing with the irradiation variable index;

[0027] Figure 4(a) and Figure 4(b) are respectively the comparison of the distribution cloud diagrams before and after optimization. Detailed Embodiments

[0028] The following is a further description of the present invention in conjunction with Figure 1 (the flow chart of the solution):

[0029] An optimization method for boron neutron capture therapy dose calculation based on a random forest model according to the present invention mainly consists of the following steps:

[0030] 1. Dose calculation and acquisition of sample data sets

[0031] In this step, by setting multiple groups of irradiation variables, the dose is calculated for each group of irradiation variables using conventional methods. By constructing a three-dimensional model of the target area, the three-dimensional model of the target area is completely described as the irradiated tissue and the region of interest; the particle transport process is calculated to give the dose values of the irradiated tissue and the region of interest; the calculation results are displayed and analyzed, and the dose calculation results under the conventional method are obtained through three steps: pre-processing, Monte Carlo calculation program, and post-processing, which are used as the sample data set.

[0032] 2. Train the random forest model.

[0033] In this step, after obtaining the results corresponding to about one hundred different irradiation variables as sample data, since the graphic data is difficult to recognize and complex, the key information of the sample data is first extracted and simplified. As shown in Table 1, for the three input variables (boron drug concentration B Dose, irradiation azimuth Angle, source strength distribution strength), they are simplified into three multi-dimensional vectors; for the output quantity, the doses corresponding to 1%, 50%, and 99% of the percentage volume of the tumor area are respectively selected as a three-dimensional vector, and another three-dimensional vector is generated for the sensitive area in the same way as above. If a patient has n tumor areas and m sensitive areas, then there are a total of (m + n) three-dimensional vectors for the output quantity. Finally, the sample data table is structured, with each row representing a group of irradiation variables and results, and all sample data is divided into an 80% model training set and a 20% model test set. In this way, the processing of the sample data is completed.

[0034] Table 1: Simplification example of sample data

[0035]

[0036] After the sample data is processed, the random forest model is built using the MATLAB program. The training set and the test set are imported, and the multi-output regression random forest model is selected. It is stipulated that 500 decision trees are generated, and each tree grows independently using different Bootstrap samples and random features. When splitting nodes, the best feature and split point are selected, and the final output form is the average of the results of all trees. After training and tuning the parameters using the test set, as shown in the appendix Figure 2 shown, the mean square error (mse) between the predicted values and the actual values of the samples is basically negligible. Therefore, this random forest model can accurately fit the sample data and make predictions. Further, the reliability verification of the random forest model prediction is carried out in step 5.

[0037] 3. Use the random forest model for prediction.

[0038] In this step, after the training is completed, the random forest model is used to predict the results of all possible irradiation variables. First, the variables corresponding to the irradiation variables to be predicted are processed into the same format as in Step 2, and the variables to be predicted are input into the MATLAB workspace dataset "data", and the corresponding data is read using the xlsread instruction. Then, the new data is defined as the input set and the same normalization conditions are applied. Finally, SVM simulation prediction is used, inverse normalization processing is performed, and the corresponding results are output. The output results are stored in the dataset "Predict" in the same data format as the data processing output in Step 2.

[0039] 4. Draw a graph of the dose effect versus different irradiation variables.

[0040] In this step, for the results of all possible irradiation variables obtained from the prediction, to facilitate quantifying the dose effect, efficiently and accurately determining the optimal irradiation variables, and reducing the loss of manpower and material resources, similar to the sample data processing in Step 2, the doses corresponding to 1%, 50%, and 99% of the percentage volume of each region are used as the basis for measuring the dose effect (in the same data format as the prediction results, and the data does not need to be processed again), and the dose effect index is calculated according to formula (1), so as to draw a graph of the dose effect versus different irradiation variables, as Figure 3 shown. The formula for calculating the dose effect index is:

[0041]

[0042] where I is the dose effect evaluation index; are the doses corresponding to 99%, 50%, and 1% of the percentage volume of the tumor region in the DVH diagram respectively; are the doses corresponding to 99%, 50%, and 1% of the percentage volume of the sensitive region in the DVH diagram respectively.

[0043] The physical meaning of this formula is: the dose effect evaluation index is directly proportional to the doses of 99%, 50%, and 1% of the percentage volume of the tumor region; and inversely proportional to the doses of 99%, 50%, and 1% of the percentage volume of the sensitive region. Considering that the doses of 99% and 1% of the percentage volume of the tumor region are more concentrated, it indicates that the irradiation of the tumor region is more uniform, and the dose effect evaluation index is also inversely proportional to the dose difference between 99% and 1% of the percentage volume of the tumor region.

[0044] 5. Verify the reliability of the random forest model.

[0045] In this step, a part of the irradiation variables is randomly selected, and the results (correct results) obtained through pre-processing, calculation, and post-processing are compared and calculated with the results predicted by the random forest model to confirm that the error is within the allowable range and the trend of the change curve is correct. If the results predicted for each group satisfy formula (2), the random forest model is reliable; if the error is large and formula (2) is not satisfied, preliminary irradiation variables are added, and after obtaining more sample data, the random forest model is retrained. After obtaining a random forest model with better prediction effect, the reliability of the random forest model is verified again. The criterion for verifying the reliability of the random forest model is:

[0046]

[0047] where x i (i = 1, 2, 3) are the doses corresponding to 1%, 50%, and 99% of the tumor region percentage volume in the DVH diagram predicted for a set of irradiation variables respectively; a i (i = 1, 2, 3) are the doses corresponding to 1%, 50%, and 99% of the tumor region percentage volume in the DVH diagram calculated for a set of irradiation variables respectively; y j (j = 1, 2, 3) are the doses corresponding to 1%, 50%, and 99% of the sensitive region percentage volume in the DVH diagram predicted for a set of irradiation variables respectively; b j (j = 1, 2, 3) are the doses corresponding to 1%, 50%, and 99% of the sensitive region percentage volume in the DVH diagram calculated for a set of irradiation variables respectively.

[0048] 6. Verify the dose optimization result.

[0049] In this step, after verifying the reliability of the random forest model in step 5, using the graph of the dose effect varying with different irradiation variables in step 4, find the irradiation variables corresponding to the highest point of the dose effect evaluation index in this graph. This set of irradiation variables is the theoretically optimal irradiation variables. Calculate the error between the dose calculation results obtained by pre-processing, Monte Carlo calculation program calculation, and post-processing of this set of theoretically optimal irradiation variables and the prediction results of the random forest model to verify the accuracy of the prediction of this set of theoretically optimal irradiation variables. On the premise of accuracy, verify that the optimization result has a real improvement through the comparison of the effects before and after optimization. As shown in Figures 4(a) and 4(b), under the condition of the same calculation example (100 groups), by comparing the cloud diagram of the irradiation variable distribution before optimization (Figure 4(a)) with the cloud diagram of the optimal irradiation variable distribution after optimization (Figure 4(b)), it can be seen that there is an obvious improvement in the dose effect after optimization.

Claims

1. An optimization method for boron neutron capture therapy dose calculation based on a random forest model, characterized in that: It includes the following steps: Step 1: Dose calculation and acquisition of sample data set: Set multiple groups of irradiation variables, each group of irradiation variables including boron drug concentration, irradiation orientation and source strength distribution. Conduct conventional dose calculation through pre-processing, Monte Carlo calculation program calculation and post-processing, and obtain the corresponding dose calculation results respectively as the sample data set; Step 2: Train the random forest model: Obtain the doses corresponding to several groups of different irradiation variables as sample data. After extracting key information and simplifying the sample data, use the sample data after extracting key information and simplifying to train the random forest model; Step 3: Use the random forest model for prediction: After training is completed, use the random forest model to predict the results of all possible irradiation variables; Step 4: Draw a graph of the dose effect index changing with different irradiation variables: From the results of all possible irradiation variables obtained by prediction, calculate the dose effect index according to formula (1) as the basis for measuring the dose effect, so as to draw a graph of the dose effect index changing with different irradiation variables; Among them, I is the dose-effect evaluation index; They are the doses corresponding to the tumor region percentage volumes of 99%, 50%, and 1% in the DVH diagram, respectively; They are the doses corresponding to the sensitive region percentage volumes of 99%, 50%, and 1% in the DVH diagram, respectively; Step 5: Verify the reliability of the random forest model: Randomly select a part of the irradiation variables, calculate the error between the dose results obtained through pre-processing, Monte Carlo calculation program calculation and post-processing and the dose results predicted by the random forest model, and confirm whether the error is within the allowable range. If it exceeds the range, repeat steps 1 - 3, train a random forest model with better prediction effect and then verify the reliability again; Step 6: Verify the dose optimization result: After verifying the reliability of the random forest model in step 5, adopt the graph of the dose effect changing with different irradiation variables in step 4, find the irradiation variables corresponding to the highest point of the dose effect evaluation index in this graph. This group of irradiation variables is the theoretically optimal irradiation variables. Calculate the error between the dose calculation results obtained by pre-processing, Monte Carlo calculation program calculation and post-processing of this group of theoretically optimal irradiation variables and the prediction results of the random forest model, verify the accuracy of the prediction of this group of theoretically optimal irradiation variables, and on the premise of accuracy, verify that the optimization result has a real improvement through the comparison of the effects before and after optimization.

2. The optimized method for boron neutron capture therapy dose calculation based on a random forest model according to claim 1, wherein: The random forest model adopts a multi-output regression random forest model.

3. An optimization method for boron neutron capture therapy dose calculation based on a random forest model according to claim 1, characterized in that: The sample data is divided into an 80% model training set and a 20% model test set.

4. An optimization method for boron neutron capture therapy dose calculation based on a random forest model according to claim 1, characterized in that: The criterion for verifying the reliability of the random forest model is: Among them, x i are the doses corresponding to the tumor region percentage volumes of 1%, 50%, and 99% in the DVH diagram obtained by predicting a set of irradiation variables respectively; a i are the doses corresponding to the tumor region percentage volumes of 1%, 50%, and 99% in the DVH diagram calculated by a set of irradiation variables respectively; y j are the doses corresponding to the sensitive region percentage volumes of 1%, 50%, and 99% in the DVH diagram obtained by predicting a set of irradiation variables respectively; b j are the doses corresponding to the sensitive region percentage volumes of 1%, 50%, and 99% in the DVH diagram calculated by a set of irradiation variables respectively.

Images