Deep Learning for Brain MRI Artifact Correction: Current Challenges and Future Directions.
Authors
Affiliations (6)
Affiliations (6)
- Department of Anatomy and Medical Imaging, Faculty of Medical and Health Sciences, University of Auckland, Auckland 1023, New Zealand.
- Auckland Bioengineering Institute, University of Auckland, Auckland 1010, New Zealand.
- Centre for Brain Research, Faculty of Medical and Health Sciences, University of Auckland, Auckland 1023, New Zealand.
- Mātai Medical Research Institute, Gisborne 4010, New Zealand.
- Medical Imaging Research Center, Faculty of Medical and Health Sciences, University of Auckland, Auckland 1023, New Zealand.
- Centre for Co-Created Ageing Research, Faculty of Medical and Health Sciences, University of Auckland, Auckland 1023, New Zealand.
Abstract
Structural magnetic resonance imaging (sMRI) is progressively used to diagnose brain diseases; however, brain sMRI scans can be easily corrupted by artifacts, e.g., motion artifacts. To remove artifacts, deep learning (DL) algorithms have been extensively studied recently. However, their performance and the challenges currently faced in clinical practice (e.g., real-world robustness, hallucination and over-smoothing) have not been adequately studied in a quantitative manner. In this structured literature review, we quantitatively examined DL-based artifact correction studies (<i>N</i> = 30), retrieved from the major databases (i.e., Google Scholar, PubMed, Web of Science, and Scopus), which particularly focused on clinical-field-strength (defined as 1.5 Tesla (T) and above) sMRI in a non-pediatric setting. Our review suggests that current DL-based approaches exhibit promising fidelity measured by structural similarity (SSIM, 0.92 ± 0.05) index and peak signal-to-noise ratio (PSNR, 32.85 ± 4.53 dB). In addition, We identified the factors underlying hallucination or over-smoothing, which are associated with neural network (NN) architecture and the training process. This study also reveals the potential advantages, brought about by frequency-aware NN. Finally, we outline several future directions, including an emerging paradigm in DL, namely physics-informed NN (PINN).