ESTRO 2023 - Abstract Book

S130

Saturday 13 May

ESTRO 2023

Materials and Methods 3152 pediatric patients from the French Childhood Cancer Survivor Study dataset who underwent 2D conventional and 3D conformational radiotherapies between 1953 and 2013 with accelerators operating with photons at energies higher than 1 MV were considered in this study (Table 1). A 3D U-Net architecture was investigated, considering in-field doses and patient geometries as inputs while using whole-body dose maps estimated with analytical models (Dos-EG) as ground truths. In addition to the traditional test set (test set), data from one specific center (center test set) and one specific linear accelerator (linac test set) were kept unseen during the neural network training. Data were split in 67%, 14%, 14%, 2% and 3% for the train, validation, test, center test and linac test set respectively.

Table 1: Distribution of accelerators and pathologies in the work cohort

Results Figure 1 shows an example of out-of-field dose prediction, and quantitative Root Mean Square Difference (RMSD) results can be found in Table 2.

Figure 1: Example of predicted out-of-field dose map (with blue square) VS ground truth out-of-field dose map, for a male patient treated on a Co60 accelerator for a nephroblastoma

RMSD

Validation

Test

Center test

Linac test

Out-of-field area

2.33e-1 cGy.Gy-1 2.38e-1 cGy.Gy-1 3.18e-1 cGy.Gy-1 1.83e-1 cGy.Gy-1

Near the field area (from isodose 5% to 0.1%) Away from field area (beyond isodose 0.1%)

2.74e-1 cGy.Gy-1 2.78e-1 cGy.Gy-1 3.34e-1 cGy.Gy-1 1.99e-1 cGy.Gy-1

0.98e-1 cGy.Gy-1 0.83e-1 cGy.Gy-1 2.01e-1 cGy.Gy-1 1.25e-1 cGy.Gy-1

Table 2: Results at epoch 1975/2000 after a 160h training

Conclusion The results suggest that the initial hypothesis is validated, i.e. it is possible to estimate the out-of-field dose from the in field dose map and the anatomy of the patient. The results on the test set, center test set, and linac test set seems to demonstrate good generalization performances, which is promising for large-scale applications on retrospective and prospective datasets and opens the door to a better understanding of dose-response relationships in the context of radiation-induced lymphopenia. This work has benefited from the grant ANR-21-RHU5-0005 within the FRANCE2030 investment plan. PD-0170 Comparing ML performances of toxicity predictive models on a large cohort of breast cancer patients M.G. Ubeira Gabellini 1 , M. Mori 1 , A. Cicchetti 1,2 , P. Mangili 1 , G. Palazzo 1 , A. Fodor 3 , A. Del Vecchio 1 , N.G. Di Muzio 4,5 , C. Fiorino 1 1 IRCCS San Raffaele Scientific Institute, Medical Physics, Milan, Italy; 2 Progetto prostata, Fondazione IRCCS Istituto Nazionale dei Tumori, Medical Physics, Milan, Italy; 3 IRCCS San Raffaele Scientific Institute, Radiotherapy, Milan, Italy; 4 IRCCS San Raffaele Scientific Institute, Radiotherapy, Milan, Italy; 5 Vita-Salute San Raffaele University, Radiotherapy, Milan, Italy Purpose or Objective Studies comparing the performances of machine learning (ML) methods in building predictive models of toxicity in RT are rare. Thanks to the availability of a large cohort (n=1323) of breast cancer patients homogeneously treated with tangential fields (Fodor et al. Clin Breast Cancer 2022), whose acute toxicity was prospectively scored, several ML approaches could be compared. Materials and Methods The endpoint was RTOG G2/G3 acute toxicity, resulting in 209 and 1114 patients respectively with or without the event. The dataset, including 25 clinical, anatomical and dosimetric features, was split into 992 for the training and 331 for external test.