ESTRO 2025 - Abstract Book

S3373

Physics - Machine learning models and clinical applications

ESTRO 2025

Results: The training/validation/testing dataset included 220/40/42 patients resulting in 79,200/14,400/15,150 samples. The AI-RTPlans were created in 0.6s/0.1s in 2D/3D single-model and 1.7min/2.1min in 2D/3D multi-model. The error between predicted and clinical meterset weight was mean±standard deviation(SD)=-0.4±3.8%/-0.23±4.8% in 2D/3D single-model and 0.01±3.9%/-0.1±5.0% in 2D/3D multi-model. The error between predicted and clinical MU per beam was mean±SD=-0.9±5.6%/-1.19±4.6% in 2D/3D single-model and 0.4±4.8%/0.5±5.3% in 2D/3D multi-model. The γPR(3%,3mm) was mean±SD=99.4±1.5%/97.6±4.4% in 2D/3D single -model and 99.9±0.3%/96.7±10.1% in 2D/3D multi-model. In both 2D models, 39/42 patients had similar or more achieved clinical objectives with the AI-RTPlan than with the clinical plan. The dosimetry of both 3D models was characterised by target underdosage (Fig. 2).

Conclusion: Accurate and fast prediction of the MU per CP with deep learning is feasible for automated VMAT treatment planning in prostate cancer.

Keywords: prostate cancer, monitor units, deep learning

References: [1] Boyer A, Biggs P, Galvin J, et al. Basic Applications of Multileaf Collimators .; 2001. doi:10.37206/71 [2] Guo C, Huang P, Li Y, Dai J. Accurate method for evaluating the duration of the entire radiotherapy process. J Appl Clin Med Phys 2020;21:252 – 8. [3] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature . 2015;521(7553):436-444. doi:10.1038/nature14539