Abstract

Background

In pencil beam scanning proton therapy, treatment is divided into small spots, consisting of the beam position and the energy, that together sculpt the radiation dose to match the tumor shape. Since protons deposit the largest dose at the end of their range, none of them emerge distal to the patient. As such, it is impossible to tell if they stopped at the intended depth.

Approach

We trained an uncertainty-aware machine learning model to predict where protons stopped (the Bragg peak) inside the patient, along with a confidence score for that prediction, based on the readout of a digital tracking calorimeter. When these predictions don’t match the treatment plan, the system rejects those spots, indicating potential problems. The rate of rejected spots across a fraction can then be used as a treatment quality metric.

Results

Our system can predict the proton stopping depth down to 1.1 mm precision (RMSE) with patient-specific training. The spot rejection rate metric can detect patient misalignments of 1 mm after on average around 1300 treated spots, which most treatment plans exceed.

Significance

This work paves the way to improve outcomes of proton therapy by utilizing a digital tracking calorimeter designed for proton computed tomography, which is planned to be positioned in-beam regardless, making range verification technology more widely available.

Key Contributions

  • Novel quality metric combining uncertainty quantification with clinical relevance
  • ML-based Bragg peak prediction with confidence intervals
  • Clinical integration pathway for real-time implementation in proton therapy
  • Robust methodology validated on anthropomorphic phantoms

Technologies & Methods

  • Python with machine learning frameworks (PyTorch)
  • Deep learning for Bragg peak prediction
  • Uncertainty quantification using Bayesian methods
  • Digital tracking calorimeter as range verification device
  • GATE/GEANT4 Monte Carlo simulation

Code and Data