KEY POINTS
- This multicentre retrospective cohort study developed BiHU-GAN using 606 paired head-and-neck non-contrast and contrast-enhanced planning CT cases, with independent external testing in 140 SegRap2023/2025 cases.
- The bidirectional model generated synthetic contrast-enhanced CT from non-contrast CT for delineation and synthetic non-contrast CT from contrast-enhanced CT for dose calculation, using CT-number deviation, gradient-consistency, and attention-based losses.
- Contours drawn on synthetic contrast-enhanced CT achieved GTV nodal Dice scores of 0.940–0.973 and HD95 values of 2.56–4.20 mm versus reference contrast-enhanced CT contours. Clinically usable ratings ranged from 85% to 96%.
- Synthetic non-contrast CT maintained a global CT-number mean absolute deviation of ≤5.5 HU, with clinically usable ratings of 89%–95% across internal and external cohorts.
- Dose recalculation on synthetic non-contrast CT produced target deviations of −1.12% to −0.88% and gamma pass rates of 96.2% at 2%/2 mm and 98.5% at 3%/3 mm. The gamma analysis, however, included only 10 internal cases.
CLINICAL TAKEAWAY
BiHU-GAN could support a single-scan head-and-neck workflow by generating contrast-like images for nodal delineation while retaining CT-number accuracy for dose calculation. The external validation and task-based testing are encouraging, but this remains technical validation rather than implementation evidence because the study was retrospective, limited to head-and-neck imaging, excluded severe target-overlapping metal artifacts, and used only 10 cases for gamma analysis.