Deep-learning synthetized 4DCT from 4DMRI of the abdominal site in carbon-ion radiotherapy

Purpose: To investigate the feasibility of deep-learning-based synthetic 4DCT (4D-sCT) generation from 4DMRI data of abdominal patients undergoing Carbon Ion Radiotherapy (CIRT). Material and methods: A 3-channel conditional Generative Adversarial Network (cGAN) was trained and tested on twenty-six patients, using paired T1-weighted 4DMRI and 4DCT volumes. 4D-sCT data were generated via the cGAN following a 3-channels segmentation approach (air, bone, soft tissue) in two scenarios: (a) 4DCT-based approach (i.e. segmentation relying on 4DCT) and (b) 4DMRI-based approach (i.e. manual segmentation on 4DMRI, to simulate a 4DMRI-only scenario). The network was first validated on a 4D computational phantom, where a ground truth dataset was available. Subsequently, the network was tested on 6 independent held-out-of-training patients. Generated volumes were evaluated with respect to the original 4DCT based on motion analysis, similarity metrics (e.g. Mean Absolute Error (MAE), Normalized Cross Coefficient (NCC)) and dosimetric criteria, by means of recalculating clinically optimized CIRT plans on the 4D-sCT. Results: For the phantom, similarity metrics were in line with literature results, while dose volume histogram values were below 0.9 %. 4DCT-based patient results demonstrated an accurate representation with respect to the original 4DCT images (MAE: 50.64-51.29 HU), while 4DMRI-only-based results yielded higher values (MAE: 81.15-90.22 HU). Gamma pass rates (3 %/3mm) were similar to 97 % for the 4DCT-based scenario, showing dosimetric consistency between the compared 4DCT and 4D-sCT dose distributions. D95% values on GTV/CTV were within clinical tolerances for the 4DMRI-only scenario. Conclusion: Deep learning-based 4D-sCT generation shows potential to support treatment planning in abdominal tumors treated with CIRT.