AI Framework Cuts Brain Imaging Scan Time in Half While Maintaining Quality

A team of researchers has developed a streamlined MRI scanning protocol that cuts brain imaging acquisition time in half by applying explainable AI techniques—demonstrating how machine learning can solve real-world constraints in medical imaging.

The Challenge with Current Brain Imaging#

Researchers studying gray matter microstructure rely on Neurite Exchange Imaging (NEXI), a biophysical diffusion MRI approach that measures how water diffuses through brain tissue compartments. This technique provides valuable insights into neurite fraction, diffusivity, and exchange time—parameters that are crucial for understanding tissue organization. However, the method demands long, multi-shell acquisitions with varying diffusion times, making it impractical for clinical and research settings. The original protocol required 27 minutes of scan time, a significant burden for patients and a bottleneck for high-throughput studies.

The XAI Optimization Strategy#

Researchers leveraged the Connectome 2.0 ultra-high gradient scanner and built an optimization framework using explainable AI (XAI) tools—specifically XGBoost, SHAP (SHapley Additive exPlanations), and Recursive Feature Elimination. They trained the system on synthetic MRI signals to identify which measurement parameters were most informative for accurate tissue characterization.

The approach identified an optimal subset of just eight features from the original 15-feature acquisition, reducing scan time to 14 minutes while preserving the quality of parameter estimates. Crucially, this AI-selected configuration converged remarkably close to the theoretical Cramér-Rao Lower Bound (CRLB)—the statistical gold standard for optimization—without requiring complex analytical calculations.

Validation and Real-World Performance#

The team tested the streamlined protocol in vivo with seven healthy participants. The XAI-optimized selection robustly reproduced microstructural parameter estimates and maintained excellent test-retest reproducibility. When compared against the full 15-feature protocol, traditional heuristic approaches ("Mid-Range" and "Corner" sampling), and theoretical benchmarks, the XAI protocol held its ground.

In particular, simpler heuristic strategies showed significant limitations. Mid-range sampling produced biased exchange time estimates because the temporal sampling lacked sufficient diversity. Corner sampling, by contrast, generated unstable intra-neurite diffusivity measurements—exhibiting five-fold higher coefficient of variation—because the approach was sensitive to scanner noise.