Purpose: To investigate the feasibility of using AI generated synthetic CT (sCT) images produced by a commercial MR simulator for proton therapy planning in the prostate.
Methods: Same day MR and CT simulation scans were acquired for 10 prostate cancer patients. For each patient, the Philips MRCAT application was used to generate sCT images based on the MR data. To normalize any discrepancy in patient positioning, the simulation CTs were deformably registered to the corresponding sCT images. Contours were propagated from the simulation CTs to both the sCT images and newly created deformed CT images (dCT). Proton PBS plans were created using the dCT images and recalculated on the sCT images. DRRs generated from each dataset were visually compared and overall dose agreement was evaluated using dose volume histograms (DVH) and line dose measurements.
Results: The sCT and dCT images differed by an average of 20 Hounsfield units (HU). Looking at individual structures, the largest differences were observed in femoral heads in which the HU decreased by an average of 64 HU in sCT images. An examination of line dose revealed the distal margin of individual lateral beams increased approximately 3mm [1-5mm] for sCT plans compared to dCT plans. However, DVH metrics in all contoured structures were nearly identical. A representative case was re-planned overriding the density of the femoral heads in the sCT to match the dCT. This resolved the range discrepancy.
Conclusion: Prostate treatment plans generated using sCT images from a commercial MR simulator showed little difference in comparison to plans based on simulation CT images. A slight overestimation of the range was noticed along beam paths that passed through the femoral heads. We anticipate that by overriding the density of these structures using a measured offset, a systematic correction can be applied to eliminate differences during planning.