Feasibility of the photon spectrum generalisation model for rapid Monte Carlo dose calculation with a deep learning-based framework

Authors：Zhu, Jinhan; Liu, Xiaowei; Chen, Lixin; Zhang, Bailin^*; Wang, Xuetao^*

Source：RADIATION PHYSICS AND CHEMISTRY, 2023, 202: 110587.

DOI：10.1016/j.radphyschem.2022.110587

• Monte Carlo algorithm
• Radiotherapy
• Deep learning
• Dose calculation

Purpose: We propose a deep learning-based strategy for the training of the radiotherapy dose calculation using limited data based on two known energy spectra within the general range of radiotherapy linear accelerators. @@@ Methods: We constructed a dose map using complete photon and electron Monte Carlo (MC) simulations with random rectangular field sizes, iso-centres, and gantry angles for pelvic computed tomography images with Mohan 4 and 24 MV spectrum photon beams. Two trained models, the virtual dose map-MC (VDM-MC) and improved VDM-MC (IVDM-MC) were tested under Mohan 4, 6, 10, 15, and 24 MV energy conditions in rectangular and intensity-modulated radiation therapy (IMRT) fields. A 3D gamma evaluation assessed the model's performance. @@@ Results: For VD-MC, the 3%/3 mm and 2%/2 mm criteria gamma pass rates were 92.58 +/- 0.87% and 85.31 +/- 1.56%, respectively, using the rectangular field test, whereas they were 97.03 +/- 0.63% and 90.97 +/- 1.46%, respectively, using the IMRT field test. For IVD-MC, the 3%/3 mm and 2%/2 mm criteria gamma pass rates were 97.86 +/- 0.35% and 92.98 +/- 0.77%, respectively, using the rectangular field test, and 98.57 +/- 0.14% and 95.11 +/- 0.33%, respectively, using the IMRT field test. @@@ Conclusion: Feasibility of a single model to achieve accurate and rapid MC dose calculations for photons with different energy spectra was preliminarily verified. This data augmentation strategy effectively generalised the scope of the model.

• Institution
  中山大学; 广州中医药大学