Machine learning for proton path tracking in proton computed tomography.

Area of Science:

• Medical Imaging
• Computational Physics
• Machine Learning

Background:

• Proton Computed Tomography (pCT) imaging is susceptible to image degradation.
• Multiple Coulomb scattering of protons causes loss in spatial resolution and quantitative integrity.
• Accurate proton path calculation is crucial for high-quality pCT reconstructions.

Purpose of the Study:

• To develop and evaluate Machine Learning (ML) models for precise proton path calculation in pCT.
• To mitigate image artifacts caused by proton scattering.
• To improve the accuracy and efficiency of proton tracking in pCT.

Main Methods:

• Utilized two ML models: a forward neural network (NN) and XGBoost.
• Trained models on synthetic data from Geant4 Monte Carlo simulations of voxelized phantoms.
• Compared ML model accuracy against a standard analytical method based on Fermi-Eyges scattering.

Main Results:

• Both NN and XGBoost models achieved high accuracy in proton path calculation, nearing theoretical limits.
• ML methods improved accuracy by 12% compared to the analytical method, particularly for large-angle scattering.
• A NN was developed to predict path calculation errors, enabling event filtering.

Conclusions:

• ML models offer an accurate and time-efficient alternative for proton tracking in pCT.
• These models can significantly enhance the spatial resolution and quantitative accuracy of reconstructed pCT images.
• The developed error prediction NN aids in improving overall image quality by filtering detrimental proton events.