A contrastive learning network for reducing metal artifacts in post-surgical knee computed tomography

Metallic knee implants can impair diagnostic accuracy in post-surgical computed tomography (CT) imaging. This study develops a novel deep neural network, contrastive learning metal artifact reduction (CLMAR), designed to reduce metal artifacts in knee CT scans following internal fixation or total knee arthroplasty. Our retrospective study encompassed knee CT images from 66 patients, reconstructed using bone and soft tissue algorithms. We generated a simulated artifact dataset (SAD) comprising 1200 artifact-free slices for quantitative analysis, employing metrics such as peak signal-to-noise ratio (PSNR), root mean square error (RMSE), and structural similarity index measure (SSIM). Additionally, a real artifact dataset (RAD) with 100 slices containing real artifacts was created for qualitative assessment through a five-point physician manual scoring system. Quantitative analysis on the SAD demonstrated that CLMAR outperformed other MAR methods across different reconstruction algorithms and display windows (bone / soft tissue), achieving superior performance in PSNR (34.72 ± 2.56 / 26.97 ± 2.26), RMSE (28.71 ± 8.18 / 18.54 ± 4.77), and SSIM (0.9592 ± 0.0252 / 0.9272 ± 0.0295), with statistically significant improvements from initial artifact images ( P < 0.001). Qualitatively, CLMAR received the highest physician rating on the RAD (3.87 ± 0.96 / 3.45 ± 0.91), indicating better artifact reduction compared to existing methods. CLMAR effectively reduces metal artifacts in post-surgical knee CT images, improving both training and inference efficiency. This enhancement increases image readability, enabling physicians to interpret post-surgical knee CT with greater accuracy. CLMAR significantly reduces metal artifacts in post-surgical knee CT scans, enhancing diagnostic accuracy and image readability. • CLMAR significantly reduces metal artifacts in post-surgical knee CT. • CLMAR outperforms other methods in quantitative and qualitative results. • CLMAR achieves superior metrics on simulated and real artifact datasets. • CLMAR shows fine generalization across different reconstruction methods. • CLMAR produces visually superior images with enhanced diagnostic quality.