Researchers boost MRI AI with smarter data filtering

A new paper investigates data curation strategies for deep learning-based accelerated MRI reconstruction. The authors assembled a large dataset of raw k-space data from 18 public sources, totaling 1.1 million images, and built a diverse evaluation set of 48 test sets covering variations in anatomy, contrast, coil count, and other factors. They proposed and studied several data filtering strategies for state-of-the-art neural networks. Their experiments show that filtering training data yields consistent, though modest, performance gains across different training set sizes and acceleration levels—with the largest benefits when the unfiltered training set contains only a small share of in-distribution data.