ESTRO 2021 Abstract Book

S1409

ESTRO 2021

Absolute Error (MAE) and Mean Error (ME) in Hounsfield Units (HU) resulted in 3D error maps for each patient of the cohort. The propagation of these maps in the common coordinate system illustrated the mean errors for the pCT generation method. To assess local errors, a paired permutation test, with the hypothesis that the intensity of the generated sample is similar to the reference, completed the study. Thus, 3D maps showing regions of significant differences between CTs and pCTs (p value < 0.05) were produced. To illustrate the methodology use, two pCT generation methods were compared: a bulk-density, and a deep- learning method based on a Generative Adversarial Network (GAN) architecture, as described in (Largent et al, 2019). Both were applied on a population of 39 patients. For each of them, a CT scan was acquired on a GE LightSpeed RT or a Toshiba Aquilion, and a T2-weighted MRI on a Siemens Skyra 3T in treatment position. Doses were computed on both reference CTs and pCTs (78 Gy in VMAT, 2 Gy per fraction).