Monte Carlo deep learning proton radiotherapy physical biological dose modeling method and system

Abstract

This invention discloses a Monte Carlo deep learning method and system for modeling physical and biological doses in proton radiotherapy, belonging to the fields of medical physics, radiotherapy, and artificial intelligence. The method first utilizes Monte Carlo simulation to generate physical dose, linear energy transfer (LET), and relative biological effect (RBE) distribution data; then, it constructs an end-to-end deep learning framework including LSTM, Transformer, and 3D CNN to achieve rapid prediction from CT images to bioequivalent doses; through multi-scale feature extraction and multi-task learning, it simultaneously predicts physical dose and personalized RBE. This invention significantly improves the speed and accuracy of dose calculation, realizes personalized RBE assessment, and provides technical support for developing safer and more efficient proton therapy protocols.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical physics, radiotherapy and artificial intelligence technology. Specifically, it relates to a dose calculation method and system for proton radiotherapy of tumors. More specifically, it relates to a calculation method and system that uses Monte Carlo simulation data to train a deep learning model to achieve personalized physical-biological dose joint modeling from patient medical images. Background Technology

[0002] Malignant tumors are a major disease that seriously threatens human health, and radiotherapy is one of its important treatment methods. Proton radiotherapy, as an advanced radiotherapy technique, utilizes the physical properties of the Bragg peak to deliver a high dose to the tumor target area while minimizing irradiation of surrounding normal tissues, thus gaining increasingly widespread clinical application. However, the clinical application of proton therapy still faces the following technical challenges:

[0003] First, there is a trade-off between efficiency and accuracy in dose calculation. Monte Carlo (MC) simulations are the reference standard for proton dose calculations. By tracking the transport processes of a large number of particles in matter, they can accurately simulate the energy deposition of protons in human tissues. However, their calculations are time-consuming; a typical clinical treatment plan usually takes several hours to complete, making it difficult to meet the needs of real-time clinical plan adjustments. While traditional analytical algorithms are faster, their accuracy decreases when dealing with areas of high tissue heterogeneity.

[0004] Second, there is uncertainty in the assessment of relative biological effectiveness (RBE). RBE is defined as the ratio of the physical dose required to produce the same biological effect to that required for a reference beam to that required for a proton beam. In current clinical practice, the proton RBE is typically set to a fixed value of 1.1. However, studies have shown that RBE is related to various factors, including the proton's linear energy transfer (LET), physical dose level, and tissue type. Particularly at the end of the proton beam's range, the LET value increases, and the actual RBE value may be higher than 1.1. Ignoring spatial variations in RBE could affect the accuracy of treatment planning.

[0005] To more accurately predict radioactive beta-reduction (RBE), researchers have proposed various models. Phenomenological models, such as the McNamara and Wedenberg models, establish empirical relationships between RBE and parameters like local effects (LET) based on fitting experimental data. These models are relatively simple to calculate but rely on specific experimental data. Mechanistic models, such as the micro-dose kinetic model (MKM) and the local effects model (LEM), are based on radiobiological mechanisms and have a more complete theoretical foundation, but they are complex and their parameters are difficult to determine. Existing methods often fail to consider the segmented characteristics of LET values, leading to insufficient accuracy in RBE assessment in low and high LET regions.

[0006] In recent years, deep learning technology has made significant progress in the field of medical image analysis. Some studies have attempted to use deep learning methods to accelerate dose calculation, but most studies have focused on physical dose prediction, with less attention paid to personalized assessment of biological effects (RBE). Furthermore, existing methods lack a framework that unifies physical dose calculation with biological effect assessment.

[0007] Therefore, a new technical solution needs to be developed that can improve the efficiency of dose calculation and enable personalized RBE assessment, especially by using a segmented model to process different LET regions, so as to provide more accurate dosimetric basis for proton therapy planning optimization. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a Monte Carlo deep learning method and system for proton radiotherapy physical and biological dose modeling. The aim is to improve the efficiency and accuracy of proton therapy dose calculation by building an end-to-end deep learning framework to integrate physical dose calculation and biological effect assessment.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] The Monte Carlo deep learning method for proton radiotherapy physical and biological dose modeling includes two phases: offline training and online application.

[0011] In the offline training phase, Monte Carlo simulation software (such as TOPAS and Geant4) was first used to simulate the proton beam transport process in tissues under different conditions, based on standard tissue models and clinical patient CT data. The three-dimensional physical dose distribution and linear energy transfer (LET) distribution were recorded during the simulation. Based on the LET distribution, a piecewise RBE model was used to calculate the corresponding RBE distribution: the Wedenberg model was used for low LET (<10 keV / μm), and the LEM model was used for high LET (≥10 keV / μm) to better match the biological effect mechanisms of different LET regions. Through data preprocessing and augmentation, a dataset for training the deep learning model was constructed.

[0012] Secondly, an end-to-end deep learning framework was constructed, combining network structures such as Long Short-Term Memory (LSTM), Transformer, and 3D Convolutional Neural Network (3D CNN). The LSTM module is used to process sequential information along the proton beam direction, capturing the depositional features of proton energy varying with depth. The Transformer module utilizes a self-attention mechanism to model the spatial correlation of dose distribution. The 3D CNN module extracts the 3D spatial features of CT images, including tissue structure and density distribution information.

[0013] During the online application phase, the system receives patient CT images, preprocesses them, and then inputs them into a trained deep learning model to quickly generate physical dose distribution, LET distribution, and RBE distribution, ultimately calculating the bioequivalent dose distribution.

[0014] The present invention also provides a system for implementing the above method, including a data interface module, a data preprocessing module, a model storage module, a calculation processing module, a result output module, and a quality assurance module.

[0015] Beneficial effects:

[0016] Compared with existing technologies, this invention has the following technical advantages: First, it improves dose calculation efficiency by reducing the calculation time to the second level through a deep learning model, meeting the needs of rapid clinical planning. Second, it enables personalized RBE assessment by using a segmented model based on the LET distribution to predict spatially varying RBE, which is more consistent with reality than fixed RBE values. Third, it establishes a combined physical-biological dosimetry framework, integrating physical dose calculation and biological effect assessment within a unified framework, reducing the accumulation of errors from step-by-step calculations. Fourth, it has good clinical applicability, providing a user-friendly interface and robust quality control, facilitating integration into existing clinical workflows. Attached Figure Description

[0017] Figure 1 This is a flowchart of the Monte Carlo deep learning proton radiotherapy physical and biological dose modeling method in an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of the network structure of the end-to-end deep learning framework in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of the invention.

[0020] Example 1: Monte Carlo deep learning method for proton radiotherapy physical and biological dose modeling.

[0021] Reference Figure 1 The method in this embodiment includes the following steps:

[0022] Step 1: Construct the training dataset:

[0023] Step 1.1: Data Collection. CT images and treatment plan data were collected from 200 patients undergoing proton therapy, including those with head and neck (60 cases), chest (50 cases), abdominal (50 cases), and pelvic (40 cases) tumors. Standard digital phantoms were also constructed, including homogeneous water phantoms and non-homogeneous tissue phantoms.

[0024] Step 1.2: Monte Carlo simulation. Proton transport simulation was performed using TOPAS 3.9. The following parameters were set for the clinical case simulation: the physical model used was QGSP_BIC_HP, the proton energy range was 70-250 MeV (intervals of 10 MeV), and the field size was 2×2cm. 2 Up to 20×20cm 2 Irradiation angle from 0° to 360° (in 30° intervals), each field simulates 10 8 Each proton supports IMPT or PBS mode. The computational grid resolution is 2×2×2 mm. 3 The output parameters include physical dose and dose-average LET.

[0025] Step 1.3: RBE Calculation. Based on the LET distribution, RBE is calculated using a piecewise model: the Wedenberg model is used for low LET (<10 keV / μm), and the LEM model is used for high LET (≥10 keV / μm). Appropriate α / β values ​​are set for different tissues.

[0026] Step 1.4: Data Preprocessing. CT values ​​were normalized to the [-1, 1] interval, doses were normalized to the prescribed dose, and LET was logarithmically transformed. Data augmentation was performed using rotation (±5°) and translation (±5 pixels). This resulted in 8000 sets of training data.

[0027] Step 2: Build a deep learning framework:

[0028] Reference Figure 2 The end-to-end framework includes:

[0029] Step 2.1: Network Architecture Design. The input is a 256×256×64 CT image. The encoder uses a 3D ResNet backbone network to extract multi-scale features. The LSTM module contains two bidirectional LSTM layers with 256 hidden units. The Transformer module contains four encoder layers, each with eight attention heads. The decoder restores spatial resolution through upsampling and skip connections. The output includes three branches: physical dose, LET, and RBE. The RBE branch applies either the Wedenberg or LEM model segmented according to the LET value.

[0030] Step 2.2: Loss Function Design. The total loss function is:

[0031] ,

[0032] in, To predict loss for physical dose, Predicting loss for LET For RBE prediction of loss, For space smoothing regularization, to These are the weighting coefficients. The initial weight values ​​are set as follows: =1.0, =0.5, =0.5, =0.1.

[0033] Step 2.3: Model Training. The Adam optimizer was used with a learning rate of 1e-4, a batch size of 4, and 150 training epochs. Training was performed on an NVIDIA V100 GPU.

[0034] Step 3: Model Validation and Application

[0035] Step 3.1: Test set validation. The model performance was evaluated using 50 independent test cases. The Gamma pass rate (3% / 3mm) for physical doses reached over 95%, and the computation time was in the second range.

[0036] Step 3.2: Clinical Application Flow. First, input the patient's CT image. After image preprocessing and standardization, input the trained model to predict the physical dose, LET, and RBE distribution (RBE is calculated in segments). Then, calculate the bioequivalent dose, and finally visualize and output the results.

[0037] Example 2: System Implementation.

[0038] The system of the present invention includes the following modules:

[0039] The data interface module supports DICOM format import and export. The preprocessing module is responsible for CT value conversion and image registration. The calculation module uses GPU-accelerated deep learning inference to support real-time calculation of segmented RBE models. The visualization module provides dose distribution display and DVH analysis functions. The quality control module implements result range checking and anomaly alarms.

[0040] The system is deployed on a workstation equipped with an NVIDIA RTX 3090, with inference speed optimized via TensorRT.

[0041] This invention achieves rapid and accurate calculation of proton therapy doses by combining the accuracy of Monte Carlo simulations with the efficiency of deep learning. In particular, it provides personalized RBE assessment, offering an effective tool for optimizing clinical proton therapy plans.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A Monte Carlo deep learning method for modeling the physical and biological dosimetry of proton radiotherapy, characterized in that, Includes the following steps: Step 1: Construct a proton dose dataset. Use Monte Carlo simulation software to simulate proton transport on standard tissue models and patient CT data to generate a dataset containing three-dimensional physical dose distribution, linear energy transfer (LET) distribution, and relative biological effect (RBE) distribution. Step 2: Establish an end-to-end framework based on deep learning, including a physical dose prediction network, an LET distribution prediction network, and an RBE calculation network. Through multi-task learning, the mapping from patient CT images to three-dimensional bioequivalent dose distribution is realized. Step 3: Design a multi-scale feature extraction mechanism, use a 3D convolutional neural network to extract spatial features of CT images, capture dose deposition features along the proton beam direction using LSTM, and use Transformer to model spatial correlation. Step 4: Construct a loss function that includes physical and biological constraints, train a deep learning model, and output the spatially varying RBE and biological equivalent dose distribution.

2. The method according to claim 1, characterized in that, Step 1 includes: Step 1.1: Collect clinical proton therapy case data and construct a standard digital phantom library including uniform phantoms, non-uniform interface phantoms, and complex structure phantoms; Step 1.2: Simulate proton transport in the 70-250 MeV energy range using the Monte Carlo simulation program. For the clinical case simulation, the following parameters were set: the physical model used was QGSP_BIC_HP, the proton energy interval was 10 MeV, and the field size was 2×2cm. 2 Up to 20×20cm 2 Irradiation angle from 0° to 360°, in 10° intervals, each field simulates 10 8 One proton, computational grid resolution 2×2×2mm 3 Record physical dose and LET distribution; Step 1.3: Based on the LET distribution, calculate the RBE distribution using a piecewise model according to the LET value: use the Wedenberg model for low LET values ​​<10 keV / μm, and use the LEM model for high LET values ​​≥10 keV / μm, and set tissue-specific α / β parameters; Step 1.4: Normalize and augment the data to construct the training dataset.

3. The method according to claim 1, characterized in that, The end-to-end framework in step 2 includes: (1) Physical dose prediction network: The 3D U-Net architecture is adopted, the LSTM module is embedded to process the sequence information, and the Transformer encoder is introduced to model the global features; (2) LET prediction network: It shares an encoder with the physical dose prediction network and outputs the LET distribution through an independent decoder; (3) RBE computation network: Input LET distribution and organization type information, select Wedenberg model or LEM model according to LET value segmentation, fuse features through attention mechanism, and output RBE distribution.

4. The method according to claim 1, characterized in that, Step 3 includes: (1) 3D ResNet is used as the feature extraction network; (2) LSTM processes 2D slice sequences in the depth direction and outputs sequence features; (3) The Transformer calculates the dependencies between features through a self-attention mechanism; (4) The output features of LSTM and Transformer are fused through a gating mechanism.

5. The method according to claim 1, characterized in that, The loss function in step 4 is: ， in, To predict loss for physical dose, Predicting loss for LET For RBE prediction of loss, For space smoothing regularization, to These are the weighting coefficients.

6. The method according to any one of claims 1-5, characterized in that, Also includes: (1) Adopt a progressive training strategy and train each prediction task in stages; (2) Use early stopping mechanism to prevent overfitting.

7. The method according to claim 1, characterized in that, In clinical applications: (1) Receive CT images in DICOM format; (2) Input the preprocessed data into the trained model; (3) Generate physical dose, LET, and RBE distributions within seconds; (4) Calculate the bioequivalent dose and visualize the output.

8. A system for implementing the method according to any one of claims 1-5, characterized in that, include: (1) Data interface module, used to receive medical image data; (2) Preprocessing module, used for image normalization processing; (3) Calculation and processing module, which includes a trained deep learning model for dose prediction; (4) Results output module, used for dose distribution visualization and data export; (5) Quality assurance module, used for result verification and anomaly detection.

9. The system according to claim 8, characterized in that, The computing module is GPU-accelerated and supports TensorRT optimization.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-5.

Images