Transformer-based rapid dose calculation for carbon ion therapy

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Online adaptive radiotherapy (OART) enhances carbon ion therapy precision through dynamic plan adjustments.
• Clinical implementation of OART is hindered by the lack of rapid and accurate dose calculation engines for online plan adaptation and verification.

Purpose of the Study:

• To develop an end-to-end workflow for per-beam carbon ion artificial intelligence (AI) dose engines, from calculation to full-plan dose evaluation.
• To compare the suitability of Transformer and Long Short-Term Memory (LSTM) sequence backbones for clinical AI dose engine applications in carbon ion therapy.

Main Methods:

• Developed a system incorporating CT-block generation, per-beam dose calculation, and dose accumulation for the Heavy Ion Medical Machine (HIMM).
• Implemented and evaluated a Transformer model (Ci-DoTA) and an LSTM model (Ci-DoLSTM) for MC-equivalent 3D per-beam dose calculation.
• Compared model performance against Pencil Beam Algorithm (PBA) and Monte Carlo (MC) methods using gamma analysis, voxel-wise errors, and runtime on head-and-neck cases.

Main Results:

• Both Ci-DoTA and Ci-DoLSTM demonstrated high per-beam agreement with MC, achieving mean pass rates of 99.85% (Ci-DoTA) and 99.06% (Ci-DoLSTM) under γ (3%, 3 mm).
• Ci-DoTA showed consistently high full-plan agreement across patients (γ (3%, 3 mm) 99.04%-99.76%), with dose differences localized near steep gradients.
• Inference times were 14.39 ms for Ci-DoTA and 11.96 ms for Ci-DoLSTM, indicating near real-time calculation capabilities.

Conclusions:

• Both AI dose engines enable near real-time carbon ion dose calculations at the pencil beam level, supporting adaptive radiotherapy workflows.
• The Transformer backbone (Ci-DoTA) offers superior accuracy, while the LSTM backbone (Ci-DoLSTM) provides faster per-beam computation.
• The study confirms the feasibility of fast dose recalculation in carbon ion therapy and highlights the influence of anatomical heterogeneity on AI model performance.