Seminar Overview
Magnetic Resonance Imaging (MRI) acquisition is fundamentally constrained by physical gradient switching and relaxation times, leading to lengthy scan durations that limit patient throughput and induce motion artifacts. This seminar presented a systematic review and meta-analysis of deep learning architectures applied to under-sampled k-space data for accelerated magnetic resonance reconstruction.
Key Findings & Architectural Paradigms
- Model Dominance: U-Net and convolutional recurrent architectures (CRNN-MRI) formed the primary deep learning baselines across benchmarked studies.
- Evaluation Metrics: Reconstruction fidelity was quantitatively evaluated across Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) against fully-sampled ground truths.
- Complex Domain Processing: Compared real-imaginary channel splitting versus complex-valued convolutions for k-space phase preservation.
- Acceleration Limits: Documented achievable scan acceleration factors ranging between 2× and 13× across cardiac, knee, and brain MRI modalities.
Primary Challenges Identified
- Out-of-Distribution Generalization: Deep neural models trained on specific coil configurations or anatomical targets exhibit quality degradation when transferred to unseen clinical hardware.
- Subtle Pathological Hallucinations: Unconstrained deep generative priors risk smoothing or hallucinating fine microvascular structures, emphasizing the necessity of physical data consistency projections.
Literature Basis
Based on the meta-analysis paper: AI-Based Reconstruction for Fast MRI – A Systematic Review and Meta-analysis (arXiv:2112.12744).