Deep learning-based metal artifact reduction in CT for head and neck radiotherapy.
Authors
Affiliations (3)
Affiliations (3)
- Department of Medical Imaging Sciences and Technologies, Faculty of Advanced Technologies in Medicine, Iran University of Medical Sciences, Tehran, Iran.
- Department of Nuclear Medicine, IKHC, Tehran University of Medical Sciences, Tehran, Iran.
- Division of Nuclear Medicine and Molecular Imaging, Geneva University Hospital, CH-1211 Geneva, Switzerland.
Abstract
Metal artifacts from dental fillings and implants compromise CT image quality, target delineation, and dose calculation accuracy in head and neck (H&N) radiotherapy. Previous reviews have largely focused on traditional metal artifact reduction methods and image quality metrics rather than radiotherapy-specific endpoints. This systematic review evaluates deep learning-based metal artifact reduction (DL-MAR) approaches for H&N radiotherapy, emphasizing technical advances and clinical dosimetric validation. Following PRISMA guidelines, PubMed, IEEE Xplore, and Scopus were systematically searched (updated April 2026) for studies assessing DL-MAR methods in H&N CT imaging and radiotherapy. Of 227 identified records, 40 studies (11 traditional MAR and 29 DL-MAR studies) met the inclusion criteria. Reported DL architectures included convolutional neural networks, generative adversarial networks, dual-domain networks, diffusion models, and task-oriented loss functions. Among studies with clinical dosimetric validation, GAN-based and task-oriented approaches consistently reported dose deviations below 3%. Gamma pass rates exceeding 95% (3%/3 mm) were reported in a subset of these studies. Dual-domain and diffusion-based methods were evaluated primarily using image quality metrics without direct radiotherapy-specific dosimetric validation. Current evidence suggests that DL-MAR may achieve clinically acceptable dosimetric accuracy while improving artifact suppression and target visualization in H&N radiotherapy planning. Task-oriented loss functions currently demonstrate promising potential for clinical translation based on available retrospective evidence. Broader clinical implementation remains limited by reliance on synthetic training data, restricted access to sinogram data, and the absence of prospective multicenter validation studies.