Learning the Physics of Particle Transport via Transformers

Oscar Pastor-Serrano, Zoltán PerkóAAAI 2022 · DOI

Abstract

Researchers have developed DoTA (Dose Transformer Algorithm), a deep learning model that predicts how proton beams deposit energy in cancer patients with unprecedented speed and precision. By framing particle physics as a sequence-modeling task, the model achieves "gold-standard" Monte Carlo accuracy while running 4,000 times faster, potentially enabling real-time adjustments to radiation therapy during a treatment session.

Introduction

In radiotherapy, the primary objective is to maximize tumor cell destruction while minimizing damage to surrounding healthy tissue. This requires precise calculation of the "dose distribution"—the spatial map of energy deposition as a particle beam traverses the patient's anatomy.

Clinicians currently face a significant trade-off between accuracy and speed. Monte Carlo (MC) simulations provide the gold standard for accuracy but are computationally prohibitive, often requiring minutes or hours. Conversely, Analytical Pencil Beam Algorithms (PBA) are fast but lose accuracy in heterogeneous regions, such as the interface between air-filled lungs and dense bone. This computational bottleneck prevents real-time adaptive radiotherapy, where beam parameters could be adjusted dynamically to account for patient movement or physiological changes.

Methodology

The authors propose DoTA, an architecture that conceptualizes the path of a particle beam through tissue as a sequence of events.

Hybrid Architecture

The model utilizes a hybrid approach combining convolutional and transformer layers to handle spatial and sequential dependencies:

  1. Convolutional Encoder: This module extracts spatial features, such as tissue density and anatomical edges, from 2D CT scan slices.
  2. Transformer Encoder: Using causal self-attention, this component routes information between spatial features and a specific "energy token." The causal constraint ensures the model respects the physical reality that a beam's state is influenced only by the tissue it has already passed through.
  3. Convolutional Decoder: This module translates the processed latent tokens back into high-resolution 2D dose distribution images.
Overview of the DoTA architecture showing the sequence-to-sequence flow from CT slices to dose distribution.

Energy Awareness and Latent Efficiency

Unlike previous models that required separate training for different beam energies, DoTA learns energy dependence as a fundamental physical property within a single model. Furthermore, by processing low-dimensional tokens in the Transformer stage, the model avoids the quadratic computational costs typically associated with high-resolution image Transformers.

Detailed view of the sub-modules: Transformer encoder, Convolutional encoder, and Convolutional decoder.

Data and Evaluation

The model was trained and validated using 3D CT scans from five lung cancer patients, a dataset chosen for its high contrast and physical complexity.

  • Ground Truth: Data was generated using MCsquare, an open-source Monte Carlo simulator, with 3 million particles per beam to ensure a low-noise gold standard.
  • Training Set: Approximately 63,000 pairs of CT blocks and dose distributions from four patients.
  • Test Set: 3,600 pairs from a fifth, unseen patient to evaluate generalization.

Results

DoTA demonstrated exceptional performance across clinical and computational metrics:

  • Accuracy: The model achieved a 99.59% gamma pass rate, significantly outperforming clinical PBAs and previous LSTM-based deep learning approaches, particularly in complex lung geometries.
  • Computational Efficiency: DoTA is 33x faster than clinical PBAs and 4,000x faster than Monte Carlo simulations.
  • Generalization: The model successfully predicted doses for unseen patients and beam energies, indicating it has successfully learned the underlying physics of particle transport rather than merely memorizing training data.

Discussion and Future Work

DoTA represents a major step toward real-time adaptive radiotherapy. However, several challenges remain:

  • Beam Modeling: Current iterations assume a simple "pencil beam." Future work will incorporate complex beam shapes and energy spectrums.
  • Particle Variety: While focused on protons, the architecture is theoretically adaptable to photons, electrons, or heavy ions.
  • Radiobiology: Integrating biological impact models to predict how different tissues react to radiation at a cellular level is a key next step.

By providing Monte Carlo-level precision in milliseconds, DoTA removes a critical barrier to the next generation of personalized cancer treatment.

Summary updated

BibTeX
@inproceedings{pastorserrano2022learning,
  author = {Oscar Pastor-Serrano and Zoltán Perkó},
  title = {{Learning the Physics of Particle Transport via Transformers}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21466},
}
Figure