CT-number-preserving bidirectional cycle GAN for contrast addition and removal in head-and-neck radiotherapy planning CT: A multicenter retrospective cohort study.

Abstract

PurposeWe propose a CT number-preserving bidirectional cycle generative adversarial network (BiHU-GAN) for two-way synthesis between non-contrast CT (NCCT) and contrast-enhanced CT (CECT) in head-and-neck radiotherapy, aiming to generate synthetic CECT (sCECT) from NCCT to improve target delineation, and synthetic NCCT (sNCCT) from CECT to enable accurate dose calculation.Methods and materialsBiHU-GAN integrates an explicit CT number deviation loss (enforcing CT number fidelity), a gradient-consistency loss (preserving structural edges), and a convolutional block attention module (CBAM)for low-contrast feature enhancement. Experiments were conducted on 606 paired head-and-neck NCCT/CECT cases (118,779 slices) from our institution and externally validated on 140 paired cases from SegRap2023/2025. We benchmarked against recent state-of-the-art methods using standard image-quality assessments and blinded 5-point clinical scoring. In addition, we assessed contouring consistency to validate sCECT for target delineation, and dose calculation accuracy to verify sNCCT for radiotherapy dose calculation.ResultsBiHU-GAN outperformed all other methods. sCECT closely matched the CECT in structural similarity and CT number fidelity, supporting reliable nodal target delineation, with high contour agreement to references. At least 85% of synthetic images were rated clinically usable (score 4 or 5 on the 5‑point clinical scoring system), reaching 96% in the internal cohort. sNCCT maintained excellent CT number accuracy (with a global mean absolute deviation of CT numbers ≤ 5.5 Hounsfield units) and high clinical acceptability, with 89%-95% clinically usable ratings. Dose calculations on sNCCT showed no significant differences from NCCT in dose-volume histogram endpoints or 3D gamma pass rates at the 2%/2 mm and 3%/3 mm criteria (gamma pass rates ≥ 95%). By contrast, direct dose calculation on CECT yielded larger systematic dose deviations and lower gamma pass rates.ConclusionsBiHU-GAN generates anatomically faithful and CT number-accurate synthetic images-improving delineation via sCECT while preserving dose accuracy via sNCCT. It avoids contrast and extra scans, reduces patient burden, and suits contrast-contraindicated or resource-limited settings. Robust multicenter performance supports broad clinical adoption.