A generative adversarial network approach to synthetic-CT creation for MRI-based radiation therapy

Tese de mestrado integrado, Engenharia Biomédica e Biofísica (Radiações em Diagnóstico e Terapia), Universidade de Lisboa, Faculdade de Ciências, 2019 ; This project presents the application of a generative adversarial network (GAN) to the creation of synthetic computed tomography (sCT) scans from volumetric T1-weighted magnetic resonance imaging (MRI), for dose calculation in MRI-based radio therapy workflows. A 3-dimensional GAN for MRI-to-CT synthesis was developed based on a 2-dimensional architecture for image-content transfer. Co-registered CT and T1 –weighted MRI scans of the head region were used for training. Tuning of the network was performed with a 7-foldcross-validation method on 42 patients. A second data set of 12 patients was used as the hold out data set for final validation. The performance of the GAN was assessed with image quality metrics, and dosimetric evaluation was performed for 33 patients by comparing dose distributions calculated on true and synthetic CT, for photon and proton therapy plans. sCT generation time was <30s per patient. The mean absolute error (MAE) between sCT and CT on the cross-validation data set was69 ± 10 HU, corresponding to a 20% decrease in error when compared to training on the original 2D GAN. Quality metric results did not differ statistically for the hold out data set (p = 0.09). Higher errors were observed for air and bone voxels, and registration errors between CT and MRI decreased performance of the algorithm. Dose deviations at the target were within 2% for the photon beams; for the proton plans, 21 patients showed dose deviations under 2%, while 12 had deviations between 2% and 8%. Pass rates (2%/ 2mm) between dose distributions were higher than 98% and 94% for photon and proton plans respectively. The results compare favorably with published algorithms and the method shows potential for MRI-guided clinical workflows. Special attention should be given when beams cross small structures and airways, and further adjustments to the algorithm should be made ...