Related Experiment Video
Updated: Jun 27, 2026
10:35
Best Current Practice for Obtaining High Quality EEG Data During Simultaneous fMRI
Published on: June 3, 2013
Deep Learning for Brain MRI Artifact Correction: Current Challenges and Future Directions
Jiangfan Yu1,2, Sibusiso Mdletshe1, Hamid Abbasi2,3
1Department of Anatomy and Medical Imaging, Faculty of Medical and Health Sciences, University of Auckland, Auckland 1023, New Zealand.
Bioengineering (Basel, Switzerland)
|June 26, 2026
Summary
Deep learning methods show promise for correcting artifacts in structural MRI scans, achieving high fidelity. Further research is needed to address challenges like hallucination and over-smoothing in clinical applications.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Structural magnetic resonance imaging (sMRI) is crucial for diagnosing brain diseases.
- Artifacts, such as motion artifacts, frequently corrupt sMRI scans, impacting diagnostic accuracy.
- Deep learning (DL) algorithms are increasingly studied for artifact correction, but their clinical performance needs quantitative evaluation.
Purpose of the Study:
- To quantitatively review deep learning-based artifact correction studies for clinical-field-strength sMRI (≥1.5T) in non-pediatric settings.
- To assess the performance and identify challenges of current DL artifact correction methods.
- To explore potential future directions, including physics-informed neural networks (PINNs).
Main Methods:
- Conducted a structured literature review of 30 DL-based artifact correction studies.
- Focused on studies using clinical-field-strength sMRI (≥1.5T) in non-pediatric populations.
- Quantitatively analyzed performance metrics like SSIM and PSNR and identified factors contributing to hallucination/over-smoothing.
Main Results:
- Current DL methods demonstrate promising artifact correction fidelity with high SSIM (0.92 ± 0.05) and PSNR (32.85 ± 4.53 dB).
- Factors influencing hallucination and over-smoothing were linked to neural network architecture and training processes.
- Frequency-aware neural networks show potential advantages for artifact correction.
Conclusions:
- DL-based artifact correction in sMRI shows significant potential for clinical applications.
- Addressing hallucination and over-smoothing through optimized NN architecture and training is critical.
- Future research should explore advanced DL paradigms like PINNs for robust artifact correction.
