Reliability‐weighted patch learning for synthetic CT reconstruction in head and neck adaptive radiotherapy: A proof‐of‐concept study
Jeong Heon Kim, Sven Olberg, Byong Su Choi, Sridhar Yaddanapudi, Xiaoying Liang, Bo Lu, Jun Tan, Jin Sung Kim, So Hyun Ahn, Keith M. Furutani, Chris J. Beltran, Justin C. ParkAbstract
Background
Cone‐beam computed tomography (CBCT) is essential for adaptive radiotherapy (ART) but is limited by poor image quality. Existing deep learning methods for synthetic CT (sCT) generation face a critical trade‐off: diffusion models offer high fidelity but are impractically slow, while fast‐inference networks often lack structural accuracy and optimization stability.
Purpose
To develop HQ‐PatchNet, a reliability‐driven deep learning framework designed to improve CBCT‐to‐sCT image quality while maintaining rapid inference speed for online ART workflows.
Methods
HQ‐PatchNet utilizes a vision transformer (ViT)‐integrated CycleGAN architecture. It introduces a reliability‐driven, SSIM‐based selective patch weighting mechanism controlled by a temperature decay schedule. This dynamically suppresses artifact‐dominated regions and focuses optimization on anatomically reliable structures. The model was evaluated on 304 paired head‐and‐neck (HN) scans against conventional networks and recent diffusion models using quantitative metrics (MAE, SSIM, NCC, PSNR), computational speed, and clinical dosimetric accuracy (3D gamma index, 2%/2 mm).
Results
HQ‐PatchNet improved image quality, reducing Mean Absolute Error (MAE) from 44.81 ± 11.24 HU to 11.85 ± 4.31 HU and achieving an SSIM of 0.960 ± 0.055, comparable to state‐of‐the‐art diffusion models. Inference required only 6.42 ± 1.71 s per volume, compared to > 15 min for diffusion approaches. The model reduced severe metal and truncation artifacts, and representative qualitative assessment suggested visually consistent sCT generation under interfractional anatomical variation. Dosimetric assessment of complex HN IMRT plans yielded a high average gamma passing rate of 99.32 ± 0.56%.
Conclusions
HQ‐PatchNet provides a computationally efficient and clinically promising approach for daily CBCT‐guided online ART by improving CBCT‐to‐sCT image quality while maintaining rapid inference speed. Further contour‐based, multi‐institutional, and multi‐site validation remains necessary to confirm its broader clinical utility.