Speaker
Description
Objective
Mixed helium-carbon beams enable online range monitoring and imaging. Both ions share the same mass-to-charge ratio, allowing simultaneous acceleration, while helium has approximately three times the range in water. However, anatomical changes beyond the carbon Bragg peak only affect the helium range. We present a deep learning approach for spot-by-spot carbon range prediction using mixed beam detector signals and planning CT information.
Methods
Mixed-beam simulations were performed using our custom GPU Monte Carlo code, Raymond, based on Geant4 (QGSP_BIC_HP, EM opt4). The dataset comprised two-field treatment plans on 17 patient CTs with Lung and Liver targets and anterior–posterior and posterior–anterior beam directions. Simulations used end-inhale and end-exhale (phases 00 and 05), representing the largest respiratory motion. In total, 882,350 mixed helium–carbon pencil beams with 10% helium were simulated in both phases.
Mixed detector images were preprocessed by a separate defragmentation model to remove carbon fragment contributions. The range prediction network combined convolutional encoders and an attention layer, using the cleaned images from phases 00 and 05, beam energy, and the phase 00 CT and carbon dose profile. It predicted the phase 05 carbon dose profile for each spot, from which R80 and its change relative to phase 00 were calculated.
Evaluation used five-fold cross-validation across patients. Mean absolute error (MAE) and root mean squared error (RMSE) were calculated per test patient and averaged separately for Lung and Liver.
Results
For Lung, the network achieved an MAE/RMSE of 2/4 mm, compared with 3.7/6.3 mm when assuming no carbon range change and 4.6/8.2 mm when assuming equal range changes for helium and carbon. For Liver, the network achieved 1.3/2.2 mm, compared with 2.2/3.7 mm and 2.1/3.2 mm under the respective assumptions. MAE reductions relative to these baselines were 48% and 58% for Lung, and 43% and 38% for Liver. Inference, including defragmentation and R80 calculation, took approximately 7 ms per beam on an NVIDIA H200.
Outlook & Discussion
Evaluation on simulated patient treatment plans demonstrated accurate carbon range prediction for both sites. Millisecond inference supports spot-by-spot assessment, highlighting the potential of mixed beams for online treatment monitoring. Future work will address clinically relevant dose changes and experimental validation.