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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning accelerates MRI reconstruction using auto-calibrated signals (ACS).
• Optimizing deep learning models for scan-specific parallel imaging reconstruction requires objective methods.
• Characterizing image quality in accelerated MRI is crucial for clinical translation.

Purpose of the Study:

• To introduce objective methods for optimizing deep learning architecture and training for scan-specific parallel MRI reconstruction.
• To propose a novel metric, the COrrelation-Based Residual Artifact Index (COBRAI), for quantifying structured residual artifacts.
• To evaluate different convolutional neural network (CNN) architectures and training strategies for 2D brain MRI.

Main Methods:

• Objective hyperparameter optimization using grid-search with K-fold cross-validation on ACS data.
• Evaluation of single-layer and three-layer residual CNNs with real and complex convolutions.
• Development and application of the COBRAI metric for artifact quantification.
• Comparison of models on FastMRI and in-house multi-contrast 2D brain MRI datasets.

Main Results:

• The grid-search strategy successfully identified optimized hyperparameters, improving image quality metrics.
• Nonlinear activation functions were found to introduce structured residual artifacts.
• A three-layer residual linear CNN with complex convolutions and fewer parameters demonstrated superior robustness and artifact reduction.
• The proposed COBRAI metric effectively quantified structured artifacts, complementing existing metrics.

Conclusions:

• Scan-specific deep learning for MRI parallel image reconstruction can be effectively optimized using objective grid-search strategies.
• The COBRAI metric provides valuable characterization of structured artifacts, aiding model selection in accelerated MRI.
• Optimized linear CNN models enable higher acceleration rates with reduced artifacts in 2D brain MRI.