Transformer‐based rapid dose calculation for carbon ion therapy

Abstract Background Online adaptive radiotherapy (OART) improves the precision of carbon ion therapy by enabling dynamic treatment plan adjustment. However, its implementation is critically constrained by the limited development of fast and accurate dose engines for both plan adjustment and online verification. Purpose To address these challenges, this study advances per‐beam carbon ion artificial intelligence (AI) dose engines toward clinical use by establishing an end‐to‐end workflow from per‐beam calculation to full‐plan accumulated dose evaluation, and by conducting a controlled comparison between Transformer and Long Short‐Term Memory (LSTM) sequence backbones to assess their relative suitability for this task. Methods The system incorporates CT‐block generation, per‐beam dose calculation, and dose accumulation and is developed for evaluation on the Heavy Ion Medical Machine (HIMM). The patient's three‐dimensional (3D) CT is decomposed into multiple CT‐blocks to represent pencil beam energy deposition. A DoTA‐style Transformer model (Ci‐DoTA) takes CT‐blocks and explicit beam energies as inputs to calculate MC‐equivalent 3D per‐beam dose, which is then weighted and summed into a full‐plan dose. An LSTM baseline model (Ci‐DoLSTM) is implemented with identical input/output definitions and training settings. The implemented models are evaluated against a Pencil Beam Algorithm (PBA) and Monte Carlo (MC) dose calculations on head‐and‐neck (H&N) cases at both per‐beam and full‐plan levels using γ analysis, voxel‐wise error metrics, and runtime. Results Ci‐DoTA and Ci‐DoLSTM achieved high per‐beam agreement with MC. Under γ (3%, 3 mm), the mean pass rates were 99.85% (Ci‐DoTA) and 99.06% (Ci‐DoLSTM). Under γ (1%, 1 mm), the gap widened, with mean pass rates of 98.01% and 94.29%, respectively. Full‐plan evaluation on all six test patients showed consistently high agreement for Ci‐DoTA, with γ (3%, 3 mm) ranging from 99.04% to 99.76% and γ (1%, 1 mm) ranging from 92.88% to 94.42%. Dose differences were spatially localized and primarily clustered near steep gradients and density interfaces, including the lateral penumbra and distal fall‐off. CT HU Gradient Intensity (CHGI) was negatively associated with γ (1%, 1 mm), with a breakpoint at 625.53 HU/mm and limited explained variance ( R 2 ≈0.30). Per‐beam inference required 14.39 ms (Ci‐DoTA) and 11.96 ms (Ci‐DoLSTM). Conclusions Both deep learning engines support near real‐time carbon ion dose calculation at the pencil beam scale. The Transformer backbone achieves higher agreement, while the LSTM backbone is faster per beam. These results support the feasibility of fast dose recalculation in carbon ion workflows, and suggest that calculation difficulty for Ci‐DoTA increases in anatomically heterogeneous CT‐blocks, although additional factors beyond CHGI remain influential.