Physics-Aware per-Beam Dose Prediction Using Distance-Corrected Conical Fluence and Multimodal Clinical Information
Abstract
Purpose
This work aims to develop a physics-aware deep learning framework for radiotherapy dose prediction that improves accuracy and clinical efficiency. The proposed Physics-Aware Multimodal UNet (PhysMM-UNet) integrates CT images, fluence maps, organ masks, and multimodal semantic clinical information to overcome limitations of image-only models and better capture beam delivery characteristics and planning intent.
Methods
PhysMM-UNet integrates beam-level delivery information, patient anatomy, and semantic clinical context for dose prediction. Distance-corrected conical (DCC) fluence maps represent beam geometry, while CT images and anatomical masks are transformed into the beam’s eye view (BEV) coordinate system to ensure geometric consistency. High-level clinical information is encoded using a large language model (LLM) and fused into the network to reflect prescription intent and planning priorities. Model training and evaluation were performed using 10 prostate and 10 lung radiotherapy plans.
Results
Qualitative comparisons using dose line profiles and dose difference maps demonstrate improved agreement with ground truth compared with UNet, Attention UNet, and C3D models. Quantitatively, gamma analysis shows superior performance of our PhysMM-UNet. For prostate cases, the proposed method achieves a gamma passing rate of 0.993 ± 0.002 under the 3 mm/3% criterion, outperforming UNet (0.947 ± 0.024), Attention UNet (0.952 ± 0.025), and C3D (0.958 ± 0.019). Under the stricter 2 mm/2% criterion, PhysMM-UNet maintains a passing rate of 0.934 ± 0.028, compared with 0.876 ± 0.037, 0.895 ± 0.031, and 0.902 ± 0.029 for the respective baseline models. For lung cases, PhysMM-UNet achieves 0.987 ± 0.007 (3 mm/3%) and 0.924 ± 0.033 (2 mm/2%), consistently outperforming all comparison models.
Conclusion
A physics-aware multimodal dose prediction network incorporating semantic clinical information was developed, demonstrating the potential to enable accurate and interpretable dose estimation for personalized radiotherapy planning.