# An Artificial Intelligence Dose Engine for Fast Carbon Ion Treatment Planning

## TL;DR

This paper introduces a fast AI tool for carbon ion therapy dose calculations, matching the accuracy of slow Monte Carlo simulations but with much faster processing.

## Contribution

The first AI-based dose engine for carbon ion therapy that predicts dose and RBE parameters with Monte Carlo accuracy at a 400× faster speed.

## Key findings

- The AI model achieved over 98% gamma pass rates for dose, α, and β predictions in heterogeneous anatomies.
- Inference was 400× faster than Monte Carlo simulations while maintaining high accuracy.
- The model's performance was stable, with mean standard deviations below 0.5% across all predictions.

## Abstract

Monte Carlo (MC) simulations provide gold-standard accuracy for carbon ion therapy dose calculations but are computationally intensive, limiting their use in adaptive workflows. Analytical pencil beam algorithms offer speed but reduced accuracy in heterogeneous tissues. This study develops the first AI-based dose engine capable of predicting relative biological effectiveness-weighted doses. Absorbed dose, α, and β parameters for optimization are calculated at MC-level accuracy with a drastically reduced computational time.

We extended the transformer-based DoTA architecture to predict absorbed dose (C-DoTA-d), α (C-DoTA-α), and β (C-DoTA-β), introducing a cross-attention mechanism for α and β to combine dose and energy inputs. The training dataset consisted of approximately 70 000 pencil beams from 187 head-and-neck patients, with ground-truth values obtained using the GPU-accelerated MC toolkit FRED. Performance was evaluated on an independent test set using gamma pass rate (1%/1 mm), depth-dose, and isodose contour Dice coefficients. MC dropout-based uncertainty analysis was performed.

Median gamma pass rates exceeded 98% for all predictions (99.76% for dose, 99.14% for α, 98.74% for β), with minima above 85% in the most heterogeneous anatomies. The Dice coefficient was 0.95 for 1% isodose contours, with slightly reduced agreement in high-gradient regions. Compared to MC FRED, inference was over 400× faster (0.032 vs 14 seconds per pencil beam) while maintaining accuracy. Uncertainty analysis showed high stability, with mean standard deviations below 0.5% for all models.

This AI-based dose engine achieves MC-quality predictions of absorbed dose and relative biological effectiveness model parameters in ∼30 ms per beamlet. Its speed and accuracy support online adaptive planning, paving the way for more effective carbon ion therapy workflows. Future work will expand to additional anatomical sites, beam geometries, and clinical beamlines.