Feasibility of deep learning-based high-quality CBCT images for adaptive radiotherapy in rectal cancer

Background Cone-beam computed tomography (CBCT) presents challenges for adaptive radiotherapy (ART) due to scatter artifacts and inaccurate CT numbers. While, deep learning-based synthetic CT (sCT) offers a potential solution, validation methods relying on deformable CT registration to assess sCT accuracy are prone to errors, particularly in the pelvic region. Therefore, a robust and clinically feasible validation approach is essential. Objective This study evaluated the feasibility of applying high-quality CBCT (HQCBCT) images, converted from CBCT using a deep learning model, for ART in rectal cancer. Methods We retrospectively evaluated rectal cancer patients who underwent radiotherapy. Original volumetric modulated arc therapy (VMAT) plans created on planning CT images were recalculated using bulk electron density assignment. Twenty patients with dose deviations < 1.5 % were selected. CBCT images were converted into HQCBCT images using a deep learning model. Multiple organs-at-risk (OARs), including the rectum, bladder, small intestine, left and right femoral heads, and spinal cord, were accurately delineated on CBCT and HQCBCT. Dose calculations on HQCBCT images were performed using native and bulk electron density values derived from the planning CT. Structural consistency was evaluated using the Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD 95 ), and volume difference (VD). Dosimetric accuracy of HQCBCT was assessed by comparing dosimetric parameters for target volumes and OARs, as well as gamma passing rates. Results For the rectum, bladder, small intestine, and spinal cord, mean DSCs exceeded 0.991, mean HD 95 s were within 0.12 mm, and mean VDs were within −0.86 %. The left and right femoral heads showed slightly lower consistency, with mean DSCs of 0.987 and 0.986, HD 95 s of 0.85 mm and 0.77 mm, and mean VDs of −1.43 % and −1.31 %, respectively. Dosimetric differences were within ± 0.15 % for target volumes and ± 0.60 % for OARs, satisfying clinical requirements. Mean gamma passing rates were 91.23 % (1 %/1 mm) and 99.37 % (2 %/2 mm). Conclusion HQCBCT images demonstrate adequate structural consistency and dosimetric accuracy, supporting their potential feasibility for ART in rectal cancer.