DuDoTransGAN: Dual-Domain transformer based generative adversarial unsupervised network for metal artifact reduction in paired and unpaired CT.

PURPOSE: Metal artifacts from implants degrade CT image quality, while supervised metal artifact reduction (MAR) methods require paired clean data that are rarely available clinically.

METHODS

In this retrospective study, we developed DuDoTransGAN, an unsupervised dual-domain generative adversarial network for paired and unpaired training. It was evaluated on a publicly available paired synthetic dataset (SYN; 90,000 training and 2,000 test slices) and a clinical unpaired knee CT dataset (KNE; 19,032 slices from 66 patients; mean age, 55.8 ± 16.1 years; 36 males), including 6,168 slices with metal artifacts. On KNE, two radiologists independently assessed five image-quality aspects using Likert scales, with ratings reviewed by a senior radiologist.

RESULTS

In the paired SYN experiment, DuDoTransGAN achieved a mean PSNR of 38.19 dB, the highest mean SSIM (0.975), and the lowest mean RMSE (5.65) among unsupervised methods. Performance remained stable with increasing metal-mask size, demonstrating robustness under larger-metal-mask conditions. In the unpaired KNE experiment, inter-rater agreement was excellent (ICC, 0.941). DuDoTransGAN achieved the highest mean scores among unsupervised methods for artifact removal (3.87 ± 0.48), diagnostic value (3.75 ± 0.50), visual effect (3.90 ± 0.36), and detail retention (3.96 ± 0.40), together with a comparable sharpness score (3.87 ± 0.58), and the highest radiologist selection rate (45 % vs 2 %-21 %).

CONCLUSION

DuDoTransGAN is an adaptable unsupervised dual-domain MAR framework that performs effectively on paired synthetic and unpaired clinical data, producing high-quality, clinically interpretable CT images in the presence of metallic implants.