A deep learning approach to generate synthetic CT in low field MR-guided adaptive radiotherapy for abdominal and pelvic cases

Purpose: Artificial intelligence (AI) can play a significant role in Magnetic Resonance guided Radiotherapy (MRgRT), especially to speed up the online adaptive workflow. The aim of this study is to set up a Deep Learning (DL) approach able to generate synthetic computed tomography (sCT) images from low field MR images in pelvis and abdomen. Methods: A conditional Generative Adversarial Network (cGAN) was used for sCT generation: a total of 120 patients treated on pelvic and abdominal sites were enrolled and divided in training (80) and test sets (40). Intensity modulated radiotherapy (IMRT) treatment plans were calculated on sCT and original CT and then compared in terms of gamma analysis and differences in Dose Volume Histogram (DVH). The two one-sided test for paired samples (TOST-P) was used to evaluate the equivalence among different DVH parameters calculated for target and organs at risks (OAR) on CT and sCT images. Results: Using a CPU architecture, the mean time required by the neural network to generate a synthetic CT was 175 ± 43 seconds (s) for pelvic cases and 110 ± 40 s for abdominal ones. Mean gamma passing rates for the three tolerance criteria analysed (1%/1 mm, 2%/2 mm and 3%/3 mm) were respectively 90.8 ± 4.5%, 98.7 ± 1.1% and 99.8 ± 0.2% for abdominal cases; 89.3 ± 4.8%, 99.0 ± 0.7% and 99.9 ± 0.2% for pelvic ones, while equivalence within 1% was observed among the DVH indicators. Conclusion: This study demonstrated that sCT generation using a DL approach is feasible for low field MR images in pelvis and abdomen, allowing a reliable calculation of IMRT plans in MRgRT.