SuperAMUS-GAN: An image-domain based supervised adversarial multi-scale U-Net synthesis model for CT metal artefacts reduction

Metal artefacts remain a major challenge in CT imaging, often obscuring critical anatomy and compromising diagnostic accuracy. While recent deep learning-based metal artefact reduction (MAR) methods have shown promising results, they often struggle with residual artefacts, lack standardised evaluations, and limiting clinical translation. We propose a supervised adversarial architecture model that integrates a multi-scale U-Net generator enhanced with dense connectivity blocks and a conditional GAN framework (SuperAMUS-GAN). It operates without the use of any prior images. The model was trained on an open-source, clinically realistic dataset including body and head CT scan images. The SuperAMUS-GAN model was evaluated with two distinct datasets: first, a detailed evaluation was performed using 1000 images to evaluate overall artefact reduction using five metrics; root mean square error (RMSE), structural similarity (SSIM), peak signal to noise ratio (PSNR), image sharpness, and bone integrity. Secondly, clinical relevance was examined using 29 hybrid clinical cases, validated and scored using three additional metrics: metal integrity, streak amplitude, and proton beam range. Results showed that the AMUS-GAN achieved an RMSE of 0.004 ± 0.002, SSIM of 0.988 ± 0.005, PSNR of 47.115 ± 2.723, image sharpness score of 1.025 ± 0.049, and bone integrity score of 0.913 ± 0.053. Additionally, in clinical scoring, AMUS-GAN attained an excellent overall score of 1.4 (on a scale from 0 to 4, where lower is better). The results show the potential of our proposed model to enhance the diagnostic ability of CT images affected by metal artefacts and support improved clinical decision-making.