Deep learning for improving in room imaging in radiotherapy: CBCT to synthetic CT conversion

The aim of this work was to test a deep learning based approach to convert Cone Beam Computed Tomography (CBCT) into Computed Tomography (CT) images for adaptive radiotherapy purposes. The algorithm was tested and validated on a cohort of 40 prostate cancer patients. A 2D U-net style network made of 5 levels was trained to remap CBCT intensities. Going deeper into the encoding part, feature image size was reduced and the number of filters was increased. The opposite happened for the decoding path. For each level, direct connections between encoding and decoding sides were present. The network was trained 3 times on 2D image pairs (CBCT as input and CT as ground truth) sliced in the axial, sagittal and coronal plane respectively. The final CT value for each voxel was obtained by running an intensity voting procedure on three candidate outputs. Conversion accuracy was quantified in terms of Mean Absolute Error (MAE) and Mean Error (ME) between sCT and CT. An average value of 31.3 ± 1.6 and -1.9 ± 4.0 Hounsfield units was obtained respectively for MAE and ME. sCT generation took 3.5 minutes on average. Future effort will be to assess accuracy from a dosimetric and tissue segmentation point of view.