Dose Distribution Prediction Using Machine Learning for Adaptive Radiotherapy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing radiotherapy methods face challenges in accurately predicting dose distribution due to anatomical changes in patients during treatment periods, leading to inefficiencies and inaccuracies in delivering radiation therapy.
Innovation Solution
A method utilizing machine learning models for dose distribution prediction, incorporating historical and current state information, and feature parameters to accurately forecast dose distribution in real-time, accounting for anatomical changes using trained models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional radiotherapy planning methods are used, then treatment plans can be created based on initial planning images, but anatomical changes during the treatment period lead to inaccurate dose distribution predictions
Solution Approach 1:
The system performs preliminary actions by pre-processing historical image data and training machine learning models before actual treatment. The dose prediction model is trained in advance on historical data, enabling rapid predictions during treatment without requiring time-consuming re-planning for each anatomical change.
Solution Approach 2:
The system creates a digital copy of the patient's anatomy through image data (CT, MRI) and uses this copy for dose distribution prediction. The machine learning model learns from historical image copies and predicts dose distribution on current image copies, avoiding the need for physical re-measurement or re-planning.
2Measurement precision
If dose distribution is recalculated for each treatment fraction based on current image data, then accuracy is improved, but computational complexity and time consumption increase significantly
Solution Approach 1:
The system replaces the traditional mechanical/computational dose calculation system with a machine learning-based prediction system. Instead of performing complex physical dose calculations for each treatment fraction, the pre-trained model predicts dose distribution directly from image data, significantly simplifying the computational process.
Solution Approach 2:
The complex dose calculation process is performed preliminarily during the model training phase. The machine learning model learns the complex relationships between anatomy and dose distribution in advance, so that during actual treatment, only simple prediction is needed rather than full dose recalculation.
3Productivity
If traditional treatment planning is performed without considering anatomical changes, then workflow is simple, but treatment effectiveness decreases due to anatomical variations
Solution Approach 1:
The system implements feedback by comparing current image data with historical image data to detect anatomical changes. The dose prediction model takes into account these changes and adjusts the predicted dose distribution accordingly, ensuring the treatment plan remains reliable despite anatomical variations during treatment.
Solution Approach 2:
The system transitions from static treatment planning (based on initial images only) to dynamic treatment planning (adapting to anatomical changes). The machine learning model enables real-time adaptation of dose predictions based on current anatomical state, making the treatment plan dynamic and responsive to changes.
Data Source
AI summary
The present disclosure provides methods and systems for dose distribution prediction. The methods comprise obtaining historical state information and current state information of a target subject. The current state information may reflect the state of the target subject in a current treatment fraction, and the historical state information may reflect the state of the target subject prior to the current treatment fraction. The methods also comprise determining a feature parameter of at least one optimization target with respect to the current treatment fraction, and predicting a current dose distribution to be used in the current treatment fraction based on at least part of the historical state information, the current state information, and the feature parameter of the at least one optimization target using a dose prediction model, the dose prediction model being a trained machine learning model.