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

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Monte Carlo (MC) simulations offer high accuracy for carbon ion therapy dose calculations but are computationally intensive.
• Analytical algorithms are faster but less accurate in heterogeneous tissues, limiting adaptive radiotherapy workflows.

Purpose of the Study:

• To develop the first AI-based dose engine for predicting relative biological effectiveness-weighted doses in carbon ion therapy.
• To achieve MC-level accuracy for absorbed dose, alpha (α), and beta (β) parameters with significantly reduced computation time.

Main Methods:

• Extended the transformer-based DoTA architecture (C-DoTA-d, C-DoTA-α, C-DoTA-β) incorporating a cross-attention mechanism.
• Trained on ~70,000 pencil beams from head-and-neck patients using MC FRED for ground truth.
• Evaluated using gamma pass rate, depth-dose, and Dice coefficients, with MC dropout for uncertainty analysis.

Main Results:

• Achieved median gamma pass rates >98% (dose: 99.76%, α: 99.14%, β: 98.74%), with minima >85% in heterogeneous regions.
• Dice coefficient for 1% isodose contours was 0.95.
• Inference speed was over 400x faster than MC (0.032s vs 14s per pencil beam) with high stability (mean SD <0.5%).

Conclusions:

• The AI dose engine provides MC-quality predictions of dose and RBE parameters rapidly (∼30 ms per beamlet).
• Its speed and accuracy facilitate online adaptive planning, enhancing carbon ion therapy.
• Future work includes expanding to diverse anatomical sites and clinical beamlines.