Generalizable Deep Learning Method for Suppressing Unseen and Multiple MRI Artifacts Using Meta-learning.

Area of Science:

• Medical imaging
• Artificial intelligence
• Signal processing

Background:

• Magnetic Resonance (MR) images are prone to artifacts from motion, resolution, and under-sampling.
• Current deep learning models often require separate training for each artifact type, limiting generalizability and efficiency.
• Joint training on multiple artifacts may not adequately address variations in artifact complexity.

Purpose of the Study:

• To introduce Curriculum-MAML (CMAML), a novel approach integrating Model-Agnostic Meta-Learning (MAML) with curriculum learning for robust MR image artifact removal.
• To enhance the adaptive learning of restoration for multiple artifacts with varying complexities during a single training process.
• To reduce the number of specialized deep learning models needed for clinical MR image artifact correction.

Main Methods:

• Implemented a nested bi-level optimization framework (MAML) to learn common knowledge across artifacts and artifact-specific restoration.
• Integrated curriculum learning into MAML (CMAML) to manage artifact complexity during training.
• Conducted comparative studies using two cardiac datasets against Stochastic Gradient Descent and standard MAML.

Main Results:

• CMAML demonstrated superior generalization, improving Peak Signal-to-Noise Ratio (PSNR) for 83% of unseen artifact types/amounts and improving Structural Similarity Index Measure (SSIM) in all cases.
• The method showed better artifact suppression for 4 out of 5 composite artifact scenarios.
• CMAML performed better in 80% of cases for images affected by multiple, combined artifacts.

Conclusions:

• CMAML effectively minimizes the need for numerous artifact-specific deep learning models, crucial for clinical deployment.
• The proposed method offers improved generalization and artifact suppression capabilities compared to existing techniques.
• CMAML presents a promising solution for enhancing the quality and reliability of clinical MR imaging.