Purpose: We investigated the geometric and dosimetric impact of 3D generative adversarial network (GAN)-based metal artifact reduction (MAR) algorithms on VMAT and IMPT for the head and neck region, based on artifact-free CT volumes with dental fillings.
Methods: Two different CT volumes, one with four (m4) and the other with eight (m8) pseudo-dental fillings with a CT value of 4,000 HU, were created as ground truths for each case from thirteen metal-free CT volumes of the head and neck region (Reference). CT volumes with metal artifacts were then generated from the Reference CT volumes (Artifacts). On the Artifacts CT volumes, metal artifacts were manually corrected for using the water density override method with a value of 1.0 g/cm3 (Water). By contrast, the CT volumes with reduced metal artifacts using 3D GAN were also generated (GAN-MAR). The structural similarity (SSIM) index within the target were calculated between the Reference CT volumes and other volumes. After creating VMAT and IMPT plans on the Reference CT volumes, the original plans were recalculated for the remaining CT volumes.
Results: The calculation time required to generate a single GAN-MAR CT volume was approximately 30 seconds. The median PTV volume was 133.6 ml (range, 98.7–169.1 ml), and median overlap ratio of the artifact corrected volume to the PTV volume was 7.5% (range, 0.0–18.5%) in the m4 group and 14.3% (range, 0.0–22.8%) in the m8 group. The SSIM showed that the GAN-MAR CT volumes were more similar to the Reference CT volumes than Water and Artifacts. The median dosimetric differences from the original plans were within 3% for VMAT and IMPT.
Conclusion: GAN-MAR yielded identical CT volumes with the Reference CT volumes. The observed dosimetric difference compared to the original plan were clinically acceptable.