ESTRO 2023 - Abstract Book

S1385

Digital Posters

ESTRO 2023

This study indicates that once decided to get started with DD implementation, the lack of clear guidelines was seen as one of the major bottlenecks. Extensive phantom measurements are planned and eight centres agreed to participate in this campaign. This will, together with the survey feedback, be the foundation for forming DD guidelines.

PO-1679 Deep learning approach to generate synthetic CT from CBCT for online adaptive radiotherapy in pelvis

L. Vellini 1 , S. Menna 2 , S. Zucca 3 , J. Lenkowicz 4 , F. Catucci 2 , F. Quaranta 2 , A. D'Aviero 2 , E. Pilloni 2 , M. Aquilano 2 , C. Di Dio 2 , M. Iezzi 2 , A. Re 2 , F. Preziosi 2 , A. Piras 5 , C. Votta 6 , D. Piccari 2 , V. Valentini 7 , L. Indovina 7 , G.C. Mattiucci 8 , D. Cusumano 2 1 University of Cagliari, Physics, Cagliari, Italy; 2 Mater Olbia Hospital, Radiation Oncology Unit, Olbia, Italy; 3 ARNAS G. Brotzu, Medical Physics, Cagliari, Italy; 4 Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Radiotherapy, Roma, Italy; 5 Villa Santa Teresa, Radiation Oncology Unit, Palermo, Italy; 6 Policlinico Gemelli, Dipartimento di Radioterapia Oncologica, Roma, Italy; 7 Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Dipartimento di Radioterapia Oncologica, Roma, Italy; 8 Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Dipartimento di Radioterapia Oncologica, Olbia, Italy Purpose or Objective Artificial intelligence (AI) is revolutionizing many aspects of radiation therapy, opening scenarios that were unimaginable just a few years ago. One of the main examples is the possibility of generating synthetic CT (sCT) images directly from the on-board imaging, so removing the CT simulation step and allowing to treat patients on the same day of the first visit, in a single-session adaptive treatment. This study aims to propose a Deep Learning (DL) approach able to fast generate sCT images from low-dose Cone Beam CT (CBCT) acquired on a modern linear accelerator integrating AI. CT images were acquired using 120 kilovoltage (kV), CBCT images with 125 kV and low-dose protocols. For the training purpose, axial CBCT and CT images were paired and anatomical correspondence in terms of air pockets, bony anatomy and skin shape was visually checked slice by slice. Paired images reporting high anatomic correspondence were selected for training. A conditional Generative Adversarial Network (cGAN) was used for sCT generation, implementing a GPU architecture with a training of 200 epochs and a batch of 1. Optimisation was carried out using Adam optimiser with a=0.999 and b=0.999 as parameters. Once completed the training, time in generating synthetic CT images from CBCT was calculated for each test case. All the procedures were run on a workstation equipped with 64 Gb RAM and 4 core (3.8GHz). The sCT image accuracy was evaluated by calculating the mean absolute error (MAE) and the mean error (ME) in terms of Hounsfield Units (HU) between the sCT and the original CT, considering the body as region of interest. Treatment plans were calculated on CT and sCT images and the percentage difference in Dose Volume Histogram (DVH) parameters was calculated to evaluate the dosimetry accuracy. Materials and Methods A total of 51 patients treated in the pelvic region were enrolled and split into training (39) and test (12). Results A total of 4647 images were selected for training and 3268 were discarded, as they do not meet the anatomical correspondence criteria. The total training time was carried out in 24 hours. An example of synthetic CT generated in the pelvic district is reported in figure 1. The mean sCT generation time was 74 seconds, with a range of 67-84 seconds. As regards the image accuracy, figure 2 reports the MAE and ME obtained for all the test cases: the average value was 36.04±6.45 HU for MAE and 10.61±4.57 HU for ME. All the DVH parameters analysed were within 1% of difference between sCT and CT.