Artifact Correction in Panoramic Radiographs Using Deep De-Shadowing

Panoramic radiography is a ubiquitous dental imaging technique that captures both jaws in a single scan. However, the 2D projection of complex 3D anatomical structures introduces prominent artifacts. These artifacts include the ghost image of the opposite jaw, the upper spine overlay, and the pharyngeal air-gap, that can obscure critical diagnostic information. In this work, we regard these artifacts as 'shadows' and introduce a novel deep learning approach for artifact correction using selective deshadowing. Our pipeline first segments each artifact using dedicated U-Net++ models, then applies ShadowFormer, a prominent transformer-based de-shadowing network, to selectively suppress both radiolucent and radiopaque artifacts. Our approach significantly improves anatomical clarity. Qualitatively, it enables clearer assessment of key diagnostic features. These include lesion texture and third molar root proximity to the mandibular canal - potentially reducing the need for CBCT. Quantitatively, it yields state-of-the-art results in mandible segmentation (Dice: 0.9764 on a public dataset) and enhances the Weber contrast in diagnostically critical regions such as the mandibular canal. By decoupling artifact interference from anatomical content, our method advances panoramic X-ray interpretation and improves diagnosis and treatment planning.