Advancements in supervised deep learning for metal artifact reduction in computed tomography: A systematic review

• Published in: European journal of radiology Vol. 181; p. 111732
• Main Authors: Kleber, Cecile E.J., Karius, Ramez, Naessens, Lucas E., Van Toledo, Coen O., A. C. van Osch, Jochen, Boomsma, Martijn F., Heemskerk, Jan W.T., van der Molen, Aart J.
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.12.2024
• Subjects: Algorithms; Deep Learning; Humans; Metals; Prostheses and Implants; Radiographic Image Interpretation, Computer-Assisted - methods; Supervised Machine Learning; Tomography, X-ray computed; Tomography, X-Ray Computed - methods; CNN; iMAR; Metals; DL; FSMAR; LBRN; DAN-Net; NMAR; FBP; IDOL-Net; DICDNet; MSE; SEMAR; ACDNet; DuDoNet; PSNR; SSIM; MARGANVAC; Tomography, X-ray computed; O-MAR; Deep Learning; Artifacts; Algorithms; MODDNet; SART; DL-MAR; InDuDoNet; LI; OSCNet; MAR
• ISSN: 0720-048X; 1872-7727; 1872-7727
• DOI: 10.1016/j.ejrad.2024.111732

[Display omitted] •Metal artefacts on CT-scans are a diagnostic problem.•Most deep learning algorithms are tested on CT scans with simulated artefacts using PSNR and SSIM.•Deep learning algorithms are a promising way to reduce metal artefacts in CT images.•Further research should be conducted using standardized metrics. Metallic artefacts caused by metal implants, are a common problem in computed tomography (CT) imaging, degrading image quality and diagnostic accuracy. With advancements in artificial intelligence, novel deep learning (DL)-based metal artefact reduction (MAR) algorithms are entering clinical practice. This systematic review provides an overview of the performance of the current supervised DL-based MAR algorithms for CT, focusing on three different domains: sinogram, image, and dual domain. A literature search was conducted in PubMed, EMBASE, Web of Science, and Scopus. Outcomes were assessed using peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) or any other objective measure comparing MAR performance to uncorrected images. After screening, fourteen studies were selected that compared DL-based MAR-algorithms with uncorrected images. MAR-algorithms were categorised into the three domains. Thirteen MAR-algorithms showed a higher PSNR and SSIM value compared to the uncorrected images and to non-DL MAR-algorithms. One study showed statistically significant better MAR performance on clinical data compared to the uncorrected images and non-DL MAR-algorithms based on Hounsfield unit calculations. DL MAR-algorithms show promising results in reducing metal artefacts, but standardised methodologies are needed to evaluate DL-based MAR-algorithms on clinical data to improve comparability between algorithms. Clinical relevance statement: Recent studies highlight the effectiveness of supervised Deep Learning-based MAR-algorithms in improving CT image quality by reducing metal artefacts in the sinogram, image and dual domain. A systematic review is needed to provide an overview of newly developed algorithms.