Purpose: Compared to photons, protons are more sensitive to organ motion and anatomy change. Verification CTs (VFCT) are often needed to evaluate the anatomy change and dose coverage during the proton treatment course. This study investigates the feasibility of using CycleGAN deep learning algorithm to generate synthetic CT (CycleGAN-sCT) from Proteus®ONE daily CBCT and use it to track the daily dose and trigger the verification CT for adaptive proton therapy.
Methods: Among 29 head and neck patients, 39 same-day CBCT-VFCT pairs from Proteus®ONE imaging system and Philips Big Bore CT were retrospectively collected. VFCT to CBCT rigid registrations were first performed to align and standardize the size of the images. They were then fed into the deep learning training algorithm as ground truth and input, respectively. sCT generated from deformable image registration (DIR-sCT) was used for comparison. The dose was calculated with the in-house commissioned MCsquare calculation engine. 3D gamma analyses were performed to evaluate the dose accuracy calculated on CycleGAN-sCT and DIR-sCT.
Results: Compared to CBCT, the preliminary results from two patients showed that both CycleGAN-sCT and DIR-sCT have noticeable improvement. The dose calculation gamma passing rate (TPS dose as reference) is increased from ~80% to ~90% for 3%/3mm, comparable to VFCT. DIR-sCT sometimes failed to deal with detailed information, while CycleGAN-sCT showed better robustness, though the 3D dose gamma passing rates from both methods are comparable to VFCT. In terms of efficiency, generate 3D CycleGAN-sCT takes a couple of seconds, and DIR-sCT takes a couple of minutes.
Conclusion: While more cases are still needed to consolidate this study, preliminary results show that the CycleGAN-sCT generated from daily CBCT for proton radiation therapy has comparable image quality as VFCT. CycleGAN-sCT can be used to track the dose and trigger the verification for adaptive proton therapy.
Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by a master research grant from Ion Beam Applications.
Not Applicable / None Entered.