Direct mapping from PET coincidence data to proton-dose and positron activity using a deep learning approach
Atiq Ur Rahman1,2, Mythra Varun Nemallapudi1, Cheng-Ying Chou3
1Institute of Physics, Academia Sinica, Taipei 11529, Taiwan.
Physics in Medicine and Biology
|August 18, 2022
Summary
Deep learning directly maps secondary particle detector data to particle therapy dose distributions. This method achieves high accuracy for intrinsic dose mapping, improving treatment planning and delivery.
Area of Science:
- Medical Physics
- Radiotherapy
- Machine Learning
Background:
- Accurate dose distribution is crucial in particle therapy.
- Secondary particle detectors offer potential for real-time dose monitoring.
- Developing algorithms for direct dose mapping from detector data is challenging.
Purpose of the Study:
- Investigate deep learning for direct mapping from detector data to intrinsic dose distributions.
- Assess the utility of conditional generative adversarial networks (cGANs) for this task.
Main Methods:
- Utilized Monte Carlo simulations (GATE/Geant4) for dataset generation.
- Trained a cGAN model on proton-irradiated CT phantom data with in-beam PET imaging.
- Evaluated model performance using mean relative error, dose fraction difference, and Bragg peak shift.
Main Results:
- Achieved <1% relative deviation in dose and <2% in range for mono-energetic beams (50-122 MeV).
- Demonstrated <1% dose and <2.6% range deviation for spread-out Bragg peaks.
- Obtained results with 10^5 coincidences acquired 5 minutes post-irradiation.
Conclusions:
- Deep learning enables direct mapping from compact detector data to dose distributions in particle therapy.
- The developed method shows promise for practical implementation and real-time monitoring.
- Future work can incorporate prior information to expand model scope and applications.
