Insights on Scan-Specific Deep-Learning Strategies for Brain MRI Parallel Imaging Reconstruction.

This study explores scan-specific deep learning strategies for parallel imaging reconstruction in brain MRI. The researchers introduced methods to objectively optimize model architecture and training details, and developed a new metric called COBRAI to quantify structured residual artifacts. Various CNN models were evaluated on FastMRI and in-house datasets, revealing that nonlinearities can produce structured artifacts. A 3-layer residual linear CNN with complex implementation and fewer parameters outperformed other models, particularly in scenarios with limited training data, enabling higher acceleration rates without compromising image quality. This work provides valuable insights for advancing deep learning-based MRI reconstruction and highlights the importance of comprehensive model evaluation, including the assessment of structured artifacts.