Virtual-simulation boosted neural network dose calculation engine for intensity-modulated radiation therapy

• Published in: Australasian physical & engineering sciences in medicine Vol. 48; no. 2; pp. 557 - 566
• Main Authors: Li, Zirong, Liu, Yaoying, Shang, Xuying, Sheng, Huashan, Xie, Chuanbin, Zhao, Wei, Zhang, Gaolong, Zhou, Qichao, Xu, Shouping
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.06.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Automation; Biological and Medical Physics; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedicine; Biophysics; Cancer therapies; Computer Simulation; Datasets; Humans; Medical and Radiation Physics; Monte Carlo Method; Neural networks; Neural Networks, Computer; Optimization techniques; Planning; Python; Radiation Dosage; Radiation therapy; Radiotherapy Dosage; Radiotherapy Planning, Computer-Assisted; Radiotherapy, Intensity-Modulated; Scientific Paper; Simulation; Workflow; Radiation therapy (RT); Neural network (NN); IMRT; Dose calculation; Virtual-simulation
• ISSN: 2662-4729; 0158-9938; 2662-4737; 2662-4737; 1879-5447
• DOI: 10.1007/s13246-025-01523-3

The Monte Carlo (MC) dose calculation method is widely recognized as the gold standard for precision in dose calculation. However, MC calculations are computationally intensive and time-consuming. This study aims to develop a neural network-based dose calculation engine using a virtual simulation database, producing dose distributions with accuracy comparable to MC dose calculations. We established an unrestricted virtual simulation database employing specific rules and automated optimization techniques. Individual dose distributions for each beam were stored. A neural network was then constructed and trained using a 3D Dense-U-Net architecture. The model’s accuracy was validated in intensity-modulated radiation therapy (IMRT) for nasopharyngeal carcinoma, cervical carcinoma, and lung cancer. A total of 31,967 single-beam doses were collected from 2,382 virtual plans. For clinical beam doses, the gamma passing rates under the 1 mm/1% and 2 mm/2% criteria improved significantly from 13.4 ± 4.8% and 37.5 ± 9.4% to 77.5 ± 7.7% and 95.6 ± 2.5%, respectively, using the model. The mean computation time was 0.017 ± 0.002 s. We successfully developed an automated training workflow for a neural network-based dose calculation model in fixed-beam IMRT. This workflow enables the generation of a substantial training dataset from a relatively small clinical dataset, resulting in a model that excels in accuracy and speed.