Accuracy study of deep learning-based synthetic CT images for adaptive radiotherapy of pelvic tumors

To evaluate the accuracy of deep learning conversion of cone-beam computed tomography (CBCT) into synthetic CT (sCT) for adaptive pelvic radiotherapy and to inform clinical application. A pelvic phantom underwent planning CT and CBCT scanning. A volumetric-modulated arc therapy plan created on the planning CT was subsequently mapped to the sCT generated from the CBCT using a deep-learning network and directly recalculated. Image quality was assessed through mean absolute error (MAE), structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and Hounsfield Unit (HU) differences for each tissue structure. Furthermore, dosimetric accuracy was examined via target and organ-at-risk dose variations together with 2 mm/2 % and 3 mm/3 % gamma passing rates. Compared with planning CT, sCT and CBCT exhibited MAEs of 5.91 and 22.20 HU, SSIMs of 0.97 and 0.85, and PSNRs of 37.46 and 25.81 dB, respectively. Notably, the largest HU deviation in sCT was 55.72 HU in the left femoral head, whereas CBCT showed 320.92 HU at the same site. Importantly, dose differences between sCT- and CT-based plans for all targets and organs at risk remained below 2.0 %, and gamma passing rates exceeded 99 % under both criteria, meeting the clinical requirements for dose accuracy and ensuring precise radiotherapy delivery. For pelvic tumors, deep-learning sCT very closely reproduces planning CT image quality and dosimetry, with dose deviations within 2 % and high gamma passing rates ensuring the safety and efficacy of radiotherapy. These findings confirm the feasibility and accuracy of implementing adaptive radiotherapy based on sCT, providing a reliable basis for its safe and efficient clinical application.