Method for rapidly predicting BNCT treatment plan by using convolutional neural network

Abstract

The invention discloses a method for rapidly predicting a BNCT treatment plan by using a convolutional neural network, and the method comprises the steps: processing an image data file and a structure set file of a specific type of tumor to generate a tensor with a specific shape, and enabling the convolutional neural network to receive the same tensor shape during training and reasoning. In the training stage, a plurality of tensors of the type are delivered to the neural network, the tensors directly predict a vector after passing through a convolution layer, a pooling layer and a full connection layer in the neural network, the vector corresponds to a specific treatment plan, then parameters of the neural network are automatically adjusted after back propagation, and the treatment plan can be quickly and accurately determined. And carrying out multi-batch and large-scale iterative training to obtain a converged neural network. After the neural network training is completed, reasoning is carried out, the reasoning speed is extremely high, the BNCT treatment plan used clinically can be directly given within the negligible time, and more choices are provided for the higher and higher precision requirement and specificity requirement for the treatment plan clinically in the follow-up process.

Description

Technical Field

[0001] The present invention relates to the field of boron neutron capture therapy (BNCT) treatment plan formulation, and in particular to a method for rapidly predicting a BNCT treatment plan using a convolutional neural network. Background Art

[0002] In the field of radiotherapy, obtaining a high-quality treatment plan can ensure that cancer cells are killed while minimizing damage to normal tissues. This is related to whether the patient's cancer can be cured and will not recur in the future. When a treatment plan is manually formulated, its quality depends on the speed of dose calculation. The faster the dose solver calculates in a limited time, the more times it can be manually iterated and optimized, and the higher the quality of the treatment plan. At present, there are two major types of methods as solvers in the field of BNCT (boron neutron capture therapy), namely deterministic methods and Monte Carlo methods. The strategy of directly calculating doses by deterministic methods is difficult to use in practical applications due to its large calculation deviation. For BNCT, since it uses a special type of neutron particles, the reaction between the particles involved and the human body will be more complicated. At present, only the Monte Carlo method can be used for solution. The Monte Carlo method has high solution accuracy, but low solution efficiency and slow speed. In the field of radiotherapy that pursues time indicators, it is generally more suitable for final calibration, but not suitable for using the Monte Carlo method to search and formulate treatment plans. A more efficient method needs to be found.

[0003] In view of the shortcomings of the current Monte Carlo method for treatment planning that cannot take into account both calculation accuracy and efficiency, a method for quickly predicting BNCT treatment plans based on convolutional neural networks was invented. Summary of the invention

[0004] A single-layer neural network with infinite neurons is used as an arbitrary function fitter. Through a large amount of training, the mapping between the input and output of any function can be accurately predicted, and a multi-layer neural network can represent any complex high-order mapping relationship to a certain extent, eliminating the complicated mathematical derivation process and having a strong adaptability at the application level. In addition, neural networks are particularly suitable for work that does not require very high numerical accuracy. The use of neural networks for rapid prediction and search of BNCT treatment plans can also avoid the disadvantage of low computational efficiency of the Monte Carlo method and ensure the quality of treatment plan formulation. Therefore, the purpose of the present invention is to provide a method for rapidly predicting BNCT treatment plans using a convolutional neural network.

[0005] In order to achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0006] A method for quickly predicting a BNCT treatment plan using a convolutional neural network comprises the following steps:

[0007] Step 1: Establish a training set for a specific type of tumor based on a large number of medical imaging data and structural set data of patients with the same case;

[0008] Step 2: Normalize the training set, including a series of numerical operations such as Hounsfield unit conversion, resolution reconstruction, material coding, and fusion coding;

[0009] Step 3: Build the required convolutional neural network, including input layer, convolution layer, pooling layer, fully connected layer and output layer;

[0010] Step 4: Based on the convolutional neural network constructed in step 3, a training sample in the training set after normalization in step 2 is delivered to the convolutional neural network to obtain the initial treatment plan prediction result;

[0011] Step 5: Use the evaluation function to evaluate and score the treatment plan prediction results of step 4;

[0012] Step 6: Use the negative number of the scoring result in step 5 as the loss function and send it back to the convolutional neural network for back propagation to optimize the parameters of each neuron in the convolutional neural network;

[0013] Step 7: For a specific training sample in step 4, use steps 4-6 to perform a large number of iterative optimizations until the loss function remains stable and maintains a small fluctuation range;

[0014] Step 8: All training samples in the training set normalized in step 2 are trained using steps 4-7. After the training, a convolutional neural network that can accurately predict the BNCT treatment plan can be obtained;

[0015] Step 9: The imaging data and structure set data of the new patient are processed using the operations in step 2. Using step 4, a high-quality BNCT treatment plan can be predicted within seconds.

[0016] Compared with the prior art, the present invention has the following outstanding advantages:

[0017] The method of the present invention abandons the traditional strategy of using a Monte Carlo solver for solving and then performing manual iterative optimization. The Monte Carlo solver is too time-consuming. In the test example, the model size is 512×512×110, containing a total of 28,835,840 grids. It takes 24 minutes to calculate and run 200 million particles using a 64-core computer. In clinical practice, there are too few times left for iterative optimization. The present invention no longer uses a Monte Carlo solver but instead uses a convolutional neural network. The convolutional neural network takes less than one millisecond to give a treatment plan, which is six orders of magnitude shorter than the traditional use of a Monte Carlo solver. It eliminates the waiting time of medical staff and physicists in the clinic, can effectively improve the efficiency of the staff, and can directly give a high-quality treatment plan in one stop, which is convenient for subsequent specialized screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a diagram of the training process for using a convolutional neural network to quickly predict BNCT treatment plans. By using the method shown in this diagram, a convolutional neural network suitable for clinical use can be trained. DETAILED DESCRIPTION

[0019] The present invention is further described in detail below through specific implementation modes.

[0020] The present invention is a method for rapidly predicting a BNCT treatment plan using a convolutional neural network, which is implemented using the following technical solutions:

[0021] Step 1: Establish a training set for a specific type of tumor based on the medical imaging data and structure set data of a large number of patients with the same case. The training set contains several patients, each of which contains medical imaging data files and structure set data files, corresponding to Figure 1 The training set in the upper left corner is composed of multiple medical images stacked along the Z direction. The number of stacks should be consistent with the number of medical images of the corresponding patient;

[0022] Step 2: Normalize the training set, including a series of numerical operations such as Hounsfield unit conversion, resolution reconstruction, material coding, and fusion coding;

[0023] The conversion of Heinz units is shown below:

[0024] Hu=pixel*slope+intercept (1)

[0025] Among them: pixel is the gray value of the i-th pixel, Hu is the CT value of the i-th pixel, slope is the scaling factor used to adjust the pixel value, which indicates the proportional relationship between the actual measured value and the original pixel value during the pixel value conversion process, intercept is a constant term used to adjust the baseline of the conversion, which indicates the expected value of the actual measured value when the original pixel value is zero, the slope value and the intercept value are all present in the medical image data file. When reconstructing the resolution, it is necessary to map the medical image data files with different numbers of pixels to the matrix shape with the same number of pixels. If the reconstructed resolution is high, the grid center point of the reconstructed resolution should be interpolated. If the coordinates of the reconstructed grid center point belong to the index of a certain grid j before reconstruction, the CT value of grid j is used. In the material coding section, the CT value is divided into 22 segments from -950 to 1500, namely -950, -120, -82, -52, -22, -8, 19, 80, 120, 200, and every 100 in the range from 200 to 1500 is a segment. The purpose of this segmentation is to encode the material. There are different materials corresponding to a specific segment, which can be used for subsequent accurate modeling. In the fusion coding section, the medical imaging data and the structure set data are merged into a high-order tensor with a total of five dimensions. The first dimension is the structure set data, which contains relevant information about organ division. Based on this dimension, it is possible to accurately determine which organ a voxel is located in. The second dimension is the material data. Based on this dimension, the specific material of a voxel can be known. The third to fifth dimensions are the index numbers of the voxel model. Based on these three dimensions, the position of a specific voxel in the voxel model can be known, corresponding to Figure 1 The 3D voxel model in the lower left corner has a resolution that is consistent with that of the medical image, which is 512×512. The picture shows a lateral slice of the human head, from which you can clearly see its constituent materials and human body structure.

[0026] Step 3: Construct the required convolutional neural network with a depth of 4, including input layer, convolution layer, pooling layer, fully connected layer, and output layer. The input layer of the convolutional neural network maintains the same shape as the tensor created in step 2. The convolution layer and the pooling layer are interlaced and coupled to extract high-dimensional features. The convolution kernel uses 3×3×3 and the step size is 1. The number of channels increases with the convolution operation. Different channels can extract high-dimensional features in different dimensions, which can expose details that are not easily perceived. The pooling layer uses the maximum pooling operation (max pool), with a size of 2×2×2 and a step size of 2. The purpose is to select the features in the convolutional tensor and select the ones with the largest feature strength. This operation will cause the original tensor size to become half of the original, but the features will be more obvious. After four convolution operations and three maximum pooling operations, the final feature model can be obtained, and then the feature model is fully connected to the output vector. The output vector is then divided into two segments of the same size. The first segment is the predicted average value of the treatment plan, and the second segment is the predicted standard deviation of the treatment plan. The two use multi-dimensional Gaussian sampling to obtain the final output vector, which is the output layer of the convolutional neural network, corresponding to Figure 1 The three-dimensional convolutional neural network in the center and the action sequence directly above it. The action sequence is the output layer vector of the three-dimensional convolutional neural network. The length of the action sequence is closely related to the subsequent mapping operation.

[0027] Step 4: Based on the convolutional neural network constructed in step 3, a training sample in the training set after normalization in step 2 is delivered to the convolutional neural network to obtain the initial treatment plan prediction result; after obtaining the output vector in step 3, it needs to be mapped to the actual clinical treatment plan, including parameters such as field angle and field weight, corresponding to Figure 1 There are different field divisions in the center and above. In the figure, a total of 2 fields are used. Each field corresponds to different variables, including angle, energy spectrum, source-skin distance and other parameters. The number of fields can be increased or decreased as needed according to different requirements. This part has high operational flexibility.

[0028] Step 5: Use the evaluation function to evaluate and score the treatment plan prediction results of step 4; submit the treatment plan prediction results to the solver for calculation. The available solvers correspond to Figure 1 The three types in the upper right corner include MCX, Phits and Hydra. You can choose any one of them. After the solver calculation is completed, a three-dimensional dose distribution can be obtained, corresponding to Figure 1 The dose distribution diagram on the right side of the center is then used to evaluate the treatment plan prediction results. The evaluation function is shown in the following formula:

[0029] evaluate=eva tumor-penalty*eva normal (2)

[0030] eva tumor =dose tumor -prescription tumor

[0031]

[0032] Among them: evaluate is the final evaluation of the quality of the treatment plan, eva tumor Evaluate the pros and cons of the treatment plan for the tumor, eva normal Evaluate the pros and cons of the treatment plan for normal tissues, dose tumor is the dose in the tumor, prescription tumor is the prescribed dose, normal[i] is the dose to a certain type of normal tissue, limit normal[i] is the dose limit of a certain type of normal tissue. Assuming there are n types of normal tissues, penalty is the penalty factor, which is generally a large positive real number. The above formula can be used to calculate the quality of the treatment plan, corresponding to Figure 1 The evaluation function in the lower right corner returns the value evaluate.

[0033] Step 6: Use the negative number of the scoring result in step 5 as the loss function and return it to the convolutional neural network for back propagation to optimize the parameters of each neuron in the convolutional neural network; the loss function is shown in the following formula:

[0034] Loss=-evaluate (3)

[0035] Using this loss function, the impact of the convolution kernel in the convolutional neural network on the final output can be evaluated and back-propagated according to the chain rule. This function can be implemented using the adaptive moment estimation algorithm (ADAM). This step corresponds to Figure 1 The process shown by the arrow directly below is to use the -evaluated value as the loss function to perform a gradient descent operation and update the three-dimensional convolutional neural network.

[0036] Step 7: For a specific training sample in step 4, use steps 4-6 to perform a large number of iterative optimizations until the loss function remains stable and is maintained in a small floating range. No specific cutoff value is set for the range of this range. Whether the training of the sample should be terminated should be determined based on the downward trend of the loss function and the average value in the stable stage. For example, for every 100 iterations, determine whether the difference between the average loss function value of these 100 iterations and the average loss function value of the previous 100 iterations is less than the difference calculated in the previous iteration, as shown in the following formula:

[0037]

[0038] If the above formula is less than zero, it means that the loss function tends to be stable and the convolutional neural network tends to converge. After making this judgment 5 times, the iterative optimization of the training sample can be stopped. This process corresponds to Figure 1 The closed loop formed by the middle and right sides.

[0039] Step 8: All training samples in the training set after normalization in step 2 are trained using steps 4-7. After the training, a convolutional neural network that can accurately predict the BNCT treatment plan can be obtained;

[0040] Step 9: Process the new patient's imaging data and structure set data using the operations in step 2. Using step 4, a high-quality BNCT treatment plan can be predicted within milliseconds. This process corresponds to Figure 1 A non-closed process formed from the upper left corner to the top.

[0041] The present invention uses a convolutional neural network to quickly predict the method of BNCT treatment plan, and generates a tensor of a specific shape by converting the image data file and the structure set file of a specific type of tumor into Heinz units, reconstructing resolution, encoding materials, fusion encoding and other steps, so that the convolutional neural network can receive the same tensor shape during training and reasoning. In the training stage, based on this type of tensor, multiple tensors of this type are delivered to the neural network, and the tensor is directly predicted to give a vector after passing through the convolution layer, the pooling layer, and the fully connected layer in the neural network. The vector corresponds to a specific treatment plan, and then an automated evaluation can be performed, and the parameters of the neural network are automatically adjusted after back propagation, and multiple batches of large-scale iterative training are performed to obtain a converged neural network. After the training of the neural network is completed, reasoning can be performed, and the reasoning speed is extremely fast, which is used to directly give a BNCT treatment plan that can be used clinically in a negligible time. Compared with the traditional manual formulation and iterative treatment plan method, it has higher efficiency, eliminates the cumbersome steps of manual adjustment, and provides more options for the subsequent clinical treatment plan with higher and higher accuracy requirements and specificity requirements.

Claims

1. A method for rapidly predicting BNCT treatment plans using a convolutional neural network, characterized in that: The steps include: Step 1: Establish a training set for a specific type of tumor based on a large number of medical imaging data and structural set data of patients with the same case; Step 2: Normalize the training set, including a series of numerical operations such as Heinz unit conversion, resolution reconstruction, material encoding, and fusion encoding; Step 3: Build the required convolutional neural network, including input layer, convolution layer, pooling layer, fully connected layer and output layer; Step 4: Based on the convolutional neural network constructed in step 3, a training sample in the training set after normalization in step 2 is delivered to the convolutional neural network to obtain the initial treatment plan prediction result; Step 5: Use the evaluation function to evaluate and score the treatment plan prediction results of step 4; Step 6: Use the negative number of the scoring result in step 5 as the loss function and send it back to the convolutional neural network for back propagation to optimize the parameters of each neuron in the convolutional neural network; Step 7: For a specific training sample in step 4, use steps 4-6 to perform a large number of iterative optimizations until the loss function remains stable; Step 8: All training samples in the training set after normalization in step 2 are trained using steps 4-7. After the training, a convolutional neural network that accurately predicts the BNCT treatment plan is obtained; Step 9: The imaging data and structure set data of the new patient are processed using the operations in step 2, and a high-quality BNCT treatment plan is predicted within seconds using step 4.

2. The method of claim 1, wherein: The evaluation function described in step 5 is as follows: evaluate=eva tumor -penalty*eva normal eca tumor =dose tumor -prescription tumor Among them: ecaluate is the final evaluation of the quality of the treatment plan, eva tumor EVA evaluates the pros and cons of treatment plans for tumors. normal Evaluate the pros and cons of the treatment plan for normal tissues, dose tumor is the dose in the tumor, prescription tumor is the prescribed dose, normal[i] is the dose to a certain type of normal tissue, limit normal[i] is the dose limit for a certain type of normal tissue. Assuming there are n types of normal tissue, penalty is the penalty factor.

3. The method of claim 1, wherein: The training set established in step 1 contains several patients, each of which contains a medical imaging data file and a structure set data file.

Images