ESTRO 2024 - Abstract Book

S3437

Physics - Dose calculation algorithms

ESTRO 2024

1 Gustave Roussy, Department of Radiation Oncology, Villejuif, France. 2 INSERM, Radiothérapie Moléculaire et Innovation Thérapeutique, Villejuif, France. 3 THERYQ, PMB-Alcen, Peynier, France

Purpose/Objective:

Radiotherapy treatments are inevitably accompanied with the delivery of low doses outside the treatment field, commonly referred to as "out-of-field dose." These doses are probably associated with the appearance of adverse effects, including radiation-induced lymphopenia and/or cancer. However, there is a lack of tools that can be used clinically to reach reliable conclusions about these dose-effects correlations based on large cohorts of patients (1). To enable the estimation of out-of-field dose maps in a clinical framework, we designed a Deep Learning (DL)-based tool for out-of-field dose estimation applicable to photon external beam radiation therapy. Experimental measurements in clinical settings were also carried out and compared to predictions of the network.

Material/Methods:

An adjusted 3D U-Net was trained for this purpose, using mean squared error as loss function and in-field dose map with binary shape of the patient whole-body anatomy as inputs. The cohort used for training and performance evaluation included 3204 paediatric patients from the French Childhood Cancer Survivor Study database, treated in 5 clinical centres before 2011 with non-modulated photon therapy techniques performed on 25 different devices (linear accelerators with high voltage > 1 MV, cobalt units and betatron devices). In this proof-of-concept, the whole body dose maps used as ground truths were computed using an empirical analytical method developed in-house (2,3). Voluntarily, the test set (486 patients) was divided into 5 data subsets, each containing patients treated with devices not seen during the learning phase, in order to assess the generalizability of the neural network on different novel radiotherapy devices. To evaluate the performance of the trained network in a clinical setting for modulated irradiation configurations and modern accelerators, plans corresponding to VMAT treatments were retrieved for 3 male patients treated for right cervical, prostate, and pelvic area, and complementary IMRT and 3D conformal radiation therapy plans were created for all and pelvic patients, respectively. All treatment plans adhered to the original clinical goals to closely replicate real-world clinical practices. All irradiation procedures were executed using a Versa HD (Elekta) operating at 6 MV. For each treatment plan, 93 radiophotoluminescent dosimeters (RPL) were strategically placed throughout the entire volume of the ATOM CIRS anthropomorphic phantom. These RPLs were previously calibrated according to a rigorous protocol (4,5). In parallel with these experimental measurements, the trained neural network was used to infer the out-of-field dose map for each of the eight treatment plans. To facilitate this, we adapted the data pre-processing pipeline to handle treatment planning system (TPS) data, requiring only the RT Dose and RT Structure in DICOM format. Only RPLs located in the out-of-field region, defined as beyond the 5% isodose, were used for network evaluation. Root-mean-square deviation (RMSD) was used as the performance metric, both at the network development and experimental stages.

Results:

The network presented average RMSD of 0.28 ± 0.08, 0.41 ± 0.26 and 0.39 ± 0.25 cGy.Gy-1 for the training, validation and test sets, respectively. Table 1 presents RMSD results comparing estimated doses with experimentally measured doses.