Deep Convolutional Neural Networks for Radiotherapy Beam Modeling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The current beam modeling process for radiation machines in radiotherapy is time-consuming and prone to errors due to measurement inaccuracies, often requiring numerous dose calculations and relying on limited datasets, which hampers the commissioning of radiation machines and patient treatment workflows.

Innovation Solution

A deep convolutional neural network is trained using synthesized datasets to determine beam model parameters, reducing the need for extensive measurement-based data and accelerating the beam modeling process by generating accurate dose profiles and parameter values.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional beam modeling with iterative dose calculations is used, then measurement accuracy can be verified, but the process is extremely time-consuming (10 minutes per calculation × 50-100 calculations)

Engineering Contradiction:
Improvedose profile measurement accuracyVSAvoidbeam modeling time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent creates synthetic copies of dose profile data through Monte Carlo simulations and measurements on phantom models. These synthetic datasets replicate the characteristics of real measurement data without requiring actual patient measurements, enabling training of neural networks that can then rapidly predict beam parameters without repeated iterative calculations

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces the mechanical iterative calculation system with a neural network-based predictive system. The neural network, trained on synthetic datasets, substitutes the iterative dose calculation process with a single-pass prediction that achieves comparable accuracy in seconds rather than hours

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Reliability

If synthesized datasets are used for training, then measurement errors are eliminated, but the datasets require extensive Monte Carlo simulations to generate

Engineering Contradiction:
Improvetraining data accuracyVSAvoiddata generation time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent performs preliminary Monte Carlo simulations and phantom measurements to generate comprehensive synthetic datasets before neural network training. This preliminary action creates a robust training foundation that eliminates the need for repeated measurements during actual beam modeling, saving time in the operational phase

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The synthetic datasets generated serve multiple functions: they train the neural network, validate the Monte Carlo engine, and provide a reusable training corpus that can be applied across different beam models and machine configurations, eliminating the need to generate new datasets for each application

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If a limited number of real radiation machine datasets are used, then measurement authenticity is maintained, but the dataset size is insufficient for effective machine learning training

Engineering Contradiction:
Improvereal measurement authenticityVSAvoidtraining dataset size
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent creates synthetic copies of real measurement data through Monte Carlo simulations that replicate the statistical properties and characteristics of actual radiation measurements. These synthetic copies expand the training dataset size while maintaining the authenticity and reliability of real measurement data

Inventive Principle:
Principle #26Copying

4Manufacturing precision

If iterative beam modeling is performed to ensure accuracy, then beam model precision is improved, but the commissioning process is delayed

Engineering Contradiction:
Improvebeam model parameter accuracyVSAvoidmachine commissioning rate
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The patent replaces the iterative mechanical adjustment process with a neural network-based predictive system. The neural network, trained on synthetic datasets, directly predicts beam model parameters from treatment planning data without requiring iterative adjustments, achieving both accuracy and speed

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The system performs self-calibration by using the neural network to automatically determine beam model parameters from treatment planning data without requiring manual iterative adjustment by physicists. The system serves itself by leveraging the trained network's predictive capabilities

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3720555B1Determining beam model parameters using deep convolutional neural networks
Publication Date: 2023.06.28 ELEKTA AB
  • EP3720555B1 patent drawingFigure 1
  • EP3720555B1 patent drawingFigure 2
  • EP3720555B1 patent drawingFigure 3

AI summary

Systems and methods can include training a deep convolutional neural network model to provide a beam model for a radiation machine, such as to deliver a radiation treatment dose to a subject. A method can include determining a range of parameter values for at least one parameter of a beam model corresponding to the radiation machine, generating a plurality of sets of beam model parameter values, wherein one or more individual sets of beam model parameter values can include a parameter value selected from the determined range of parameter values, providing a plurality of corresponding dose profiles respectively corresponding to respective individual sets beam model parameter values in the plurality of sets of beam model parameter values, and training the neural network model using the plurality of beam models and the corresponding dose profiles.