Semi-Supervised Transformer For Maxillofacial Ct Metal Artifact Reduction

This study proposes Semi-SFTrans, a semi-supervised spatial-frequency transformer, for metal artifact reduction (MAR) in maxillofacial CT, evaluated against supervised models (Sup-SFTrans, ResUNet) and linear interpolation (LI). Supervised models were trained on paired simulated CT data, while Semi-SFTrans incorporated both simulated and unpaired clinical CT data. Evaluations included simulated data metrics (PSNR/SSIM) and clinical assessments from 22 cases with intraoral scans. Three radiologists scored artifact severity (Likert scale) on original, Sup-SFTrans-, and Semi-SFTrans-corrected images. Quantitative MAR performance was analyzed via HU values, noise (SD/SNR) in soft tissue ROIs, and teeth integrity using Dice/Hausdorff metrics against intraoral scan segmentations. Results showed Sup-SFTrans achieved superior PSNR/SSIM on simulated data (vs. LI/ResUNet/Semi-SFTrans). However, Semi-SFTrans outperformed Sup-SFTrans clinically, with radiologists’ scores of 4.46±0.51 vs. 3.64±0.90 (p<0.001; ICC>0.8). Semi-SFTrans significantly reduced artifacts in all soft tissue ROIs (tongue: p<0.001; lips/buccal/masseter: p<0.05) and demonstrated higher SNR improvements (tongue: p=0.039; buccal/lips/masseter: p<0.05) versus Sup-SFTrans. Notably, Semi-SFTrans better preserved teeth structures (Dice: p<0.001; Hausdorff: p=0.0022). The semi-supervised approach effectively leverages unpaired clinical data, achieving superior artifact suppression and anatomical fidelity in real-world CT scans compared to supervised models. This work highlights the potential of semi-supervised domain adaptation for MAR, particularly when clinical paired training data are limited.