Self-learning GAN based synthetic CT generation: unlocking CBCT-based adaptive radiotherapy.

Purpose/objectives: This study proposes and clinically evaluates synthetic CT (sCT) images generated from multi-center CBCT scans using artificial intelligence, with the aim of fully leveraging CBCT for adaptive radiotherapy in patients with pelvic, head-and-neck, lung, and breast cancer. Materials and methods: In collaboration with TheraPanacea (Paris, France), AI-based sCT models were developed for multiple anatomical sites using a cycleGAN architecture. The study included 51 patients from two European institutions diagnosed with head-and-neck, lung, pelvic or breast cancer and treated with CBCT-based position verification. CBCT scans were acquired using two linear accelerator systems (Varian and Elekta). Image accuracy was assessed using MAE, SSIM, and PSNR. For dosimetric evaluation, planning CTs (pCTs) were non-rigidly registered to CBCTs. Treatment plans were created on the pCT using a clinical TPS to meet standard clinical criteria, then recalculated on both the warped CT (wCT) and sCT. Dose distributions were compared using global gamma passing rates and dose-volume metrics. Results: The proposed model substantially improved image quality compared with CBCT. MAE decreased from 122.95 ± 50.07 to 23.65 ± 10.09, while SSIM increased from 0.78 ± 0.12 to 0.97 ± 0.03 and PSNR from 35.01 ± 7.24 to 44.35 ± 7.07. Dose-metric comparisons showed strong agreement between the pCT and wCT, with median relative differences within 0.5% for both targets and organs at risk. Median gamma passing rates for 2%/2 mm and 3%/3 mm criteria (10% threshold) reached 100% across all anatomical sites. No performance differences were observed between Elekta- and Varian-sCTs. Conclusion: This multi-center study demonstrates the feasibility of generating clinically acceptable AI-based sCTs from CBCT for multiple anatomical sites, yielding consistent image quality improvements and reliable dosimetric accuracy.