ESTRO 2024 - Abstract Book

S3710

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

References:

[1] M. Ronga, U. Deut, A. Bonfrate, L. De Marzi, "Very high-energy electron dose calculation using multiple scattering theory and a simplified pencil beam model", Medical Physics, pp. 1-14 (2023).

[2] L. Whitmore, R.I. Mackay, M. van Herk, J.K. Jones, R.M. Jones, "Focused VHEE (very high energy electron) beams and dose delivery for radiotherapy applications", Scientific Reports 11, 14013 (2021).

2772

Poster Discussion

Fast and accurate dose calculation algorithm based on multi-level U-Net for IMRT

Zirong Li 1 , Yaoying Liu 2 , Xuying Shang 2 , Wei Zhao 2 , Gaolong Zhang 2 , Huashan Sheng 1 , Qichao Zhou 1 , Shouping Xu 3

1 Manteia, research and algorithms, Xiamen, China. 2 Beihang University, School of Physics, Beijing, China. 3 Cancer Hospital, Chinese Academy of Medical Sciences, Department of radiotherapy, Beijing, China

Purpose/Objective:

Radiotherapy is one of the important means of treating cancer patients. The most commonly used technology is intensity-modulated radiotherapy (IMRT). IMRT treatment plans are generally obtained by reverse optimization in the treatment planning system (TPS). The speed and quality of optimization often depend on the dose calculation algorithm. Monte Carlo (MC) is a precise algorithm for dose calculation. However, the MC algorithm takes much work to be completed fast because it needs to simulate the real transport process of a large number of particles. In addition to the MC algorithm, dose calculations based on analytical algorithms are usually high-speed, but these algorithms sacrifice high accuracy. For example, the pencil beam (PB) algorithm will be too smooth when calculating cavity dose. There are currently many studies on dose prediction based on patient CT. Although most of the results have achieved high accuracy, they do not take into account the different number and angle of beams and the shape of multi-leaf collimators (MLC) in different plans, this makes it impossible to apply these results in plan optimization. Many studies used deep learning (DL) models to accelerate MC speeds, but the model’s accuracy and generalization were limited by the finite clinical dataset. However, clinical data are valuable and scarce, which makes it difficult for models trained based on clinical data to achieve a high degree of generalization. Generally, these trained models can only be used for the cancer types and fixed planning methods included in the training data. A model with generalization capabilities is essential for applying plan optimization. Thus, we aim to break the limitation of the training dataset size and, establish a wide-generalization DL-based MC frame (gdMC).

Material/Methods:

Our training data comes from an automated optimization process, plans with different beam angles are automatically generated based on the patient's target area, and dose maps are generated using MC and PB dose calculation algorithms respectively. This eliminates the need to use real patient treatment plans. Specifically, the optimization process is divided into three steps. First, setting the Iso-center at the radiotherapy planning target volume (PTV) geometric center and the number of beams to 7. The angle of each beam is randomly set. It results in 128 random plans for each case. The second step is to optimize all plans. The optimization constraints of the target area are set to the prescription dose (D95≥70Gy) and with a random weight range from 150 to 200. on the