Authors
Soyoung Chung, Jin Ho Kim, Min Gyo Chung, Hyeongmin Jin
Published in
Journal of imaging informatics in medicine. Aug 11, 2026. Epub Aug 11, 2026.
Abstract
Limited longitudinal field of view in cone-beam computed tomography (CBCT) remains a significant challenge for image-guided adaptive radiotherapy. We developed a deep learning-based CBCT outpainting framework that synthesizes anatomically consistent extensions beyond the scanned volume. The model employs a dual-branch encoder-decoder architecture that separately encodes CBCT and edge maps derived from planning CT (pCT) and integrates them via attention-guided skip connections and late fusion to preserve structure while accommodating cross-modality differences. The framework was trained and validated using 90 prostate cancer cases, with 20 held-out patients used for quantitative evaluation against four baseline models. The proposed method achieved the lowest absolute volume difference for the clinical target volume at 0.11 ± 0.07 cc and for the planning target volume at 0.05 ± 0.04 cc. It also yielded superior perceptual image quality, with LPIPS of 0.119 ± 0.029 and FID of 54.70. Qualitative assessment confirmed improved soft-tissue continuity and more faithful preservation of anatomical landmarks in both bony and low-contrast regions. These results demonstrate that the proposed framework can restore CBCT longitudinal coverage with high anatomical fidelity, supporting more reliable deformable registration and adaptive radiotherapy workflows in clinical environments constrained by limited-FOV CBCT.
PMID:
42581241
Bibliographic data and abstract were imported from PubMed on 12 Aug 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 2
- Comments 0