Deep Learning-Based Prediction of Monte Carlo Dose Distributions from Pencil Beam Algorithms for Carbon Ion Radiotherapy
Abstract
Purpose
This study aims to develop a deep learning (DL) model capable of predicting Monte Carlo (MC)-level carbon-ion dose distributions from Pencil Beam Algorithm (PBA) calculations. The objective is to achieve the high accuracy of MC simulations while maintaining the computational efficiency of analytical models for carbon-ion radiotherapy (CIRT).
Methods
A DL model utilizing a 3D Hierarchically Densely Connected U-Net architecture was trained and validated using data from 30 patients, including prostate, lung, and liver cases. The model inputs were 3D dose distributions calculated by a commercial treatment planning system, and the corresponding ground truth distributions were generated using the OpenTOPAS MC code. Evaluation metrics included the gamma passing rate (GPR) with 2 mm/2% criteria and the structural similarity index measure (SSIM). The performance of the DL-predicted doses was assessed by comparing them against the MC ground truth, specifically focusing on the accuracy gains over the original PBA calculations.
Results
The DL-predicted doses achieved a mean GPR of 97.28% ± 0.50% against the MC ground truth, outperforming the original PBA calculations (GPR: 93.01% ± 0.30%). Structural agreement also showed improvement, with the mean SSIM increasing from 0.88 ± 0.00 for the original PBA input to 0.90 ± 0.00 for the DL output. The results show that the developed model effectively reduced dose discrepancies present in the PBA calculations, particularly in the entrance plateau and fragmentation tail regions. The computational time was reduced from 20.4 hours required for full MC simulations to less than 1 second using the DL model.
Conclusion
The developed DL model successfully bridges the gap between the computational speed of PBA and the accuracy of MC simulations. Future studies will focus on extending validation to larger cohorts and additional complex anatomical sites, such as head and neck and pancreatic cancers.