ESTRO 2024 - Abstract Book

S3877

Physics - Image acquisition and processing

ESTRO 2024

1868

Digital Poster

Realistic computational phantoms for the validation of synthetic CT of the abdomen

Francesca Camagni 1 , Anestis Nakas 1 , Giovanni Parrella 1 , Alessandro Vai 2 , Silvia Molinelli 2 , Viviana Vitolo 3 , Amelia Barcellini 3 , Sara Imparato 3 , Andrea Pella 4 , Ester Orlandi 3 , Guido Baroni 1,4 , Marco Riboldi 5 , Chiara Paganelli 1 1 Politecnico di Milano, Department of Electronics, Information and Bioengineering, Milano, Italy. 2 National Center for Oncological Hadrontherapy, CNAO, Medical Physics unit, Pavia, Italy. 3 National Center for Oncological Hadrontherapy, CNAO, Clinical unit, Pavia, Italy. 4 National Center for Oncological Hadrontherapy, CNAO, Bioengineering unit, Pavia, Italy. 5 LMU Munich, Department of Medical Physics, Faculty of Physics, Munich, Germany

Purpose/Objective:

To validate the performance of a well-trained 3-channel conditional GAN (cGAN) for MRI-based abdominal sCT generation [1] using computational CT/MRI phantoms that ensure the presence of real ground truths (GT). MRI can be adopted either offline or online for adaptive radiotherapy through the generation of synthetic CT (sCT). Deep learning (DL) methods can derive sCT from MRI data, however, they lack of real validation of their performance because of the absence of a proper GT, especially for anatomical districts affected by organ motion such as the abdomen. To properly assess the performance of a DL network, the sCT derived from MRI needs to be compared with a reference CT of GT. Organ motion in the abdomen precludes a proper matching of the anatomy between CT and MRI data and deformable image registration (DIR) usually leads to sub-optimal results when the anatomy is significantly different among acquisitions. We hypothesised that digital CT/MRI phantoms could be an appropriate validation tool. Since phantom images differ from real patient images, which were seen by the net during the training phase in terms of intensity distribution, we first improved CT/MRI digital phantoms making them close to real patients’ data through Cycle-consistent GAN networks (CycleGAN). We trained separately two CycleGANs: a CycleGAN CT to generate realistic CTs starting from XCAT CT phantoms and a CycleGAN MRI to generate realistic MRIs from the MRI version of the XCAT phantom. The networks were trained, respectively, with 20 unpaired abdominal patient and phantom CT volumes and 20 unpaired abdominal patient and phantom MRIs for 100 epochs. The realistic phantoms were built to preserve as much as possible the original phantom anatomy. Five additional CT/MRI phantoms were used for testing CycleGANs performance: realistic images were evaluated in terms of (i) anatomical accuracy for the corresponding original phantom images and in terms of (ii) realistic properties with respect to unpaired real patient images (i.e., noise magnitude and intensity distribution histograms). In (i) we calculated the Structure Similarity Index (SSIM), Feature Similarity Index (FSIM), Edge Preservation Ratio (EPR), Edge Generation Ratio (EGR) and Mean Absolute Error (MAE) on corresponding axial slices. In (ii) we computed the noise magnitude (NM) of segmented realistic and real livers as the standard deviation of the liver voxel intensities and the Pearson correlation coefficient between average realistic and patient volumetric histograms (HistCC). The realistic and co-registered CT/MRI phantom images were then used for validating MRI-based sCT generation: the realistic phantom MRIs were provided in input to the in-house 3-channel (bones, air, soft tissues) cGAN to predict the corresponding sCTs. The quality of the predicted sCT Material/Methods: