Abstract
Quality control of MR images is essential for excluding problematic acquisitions and avoiding bias in subsequent image processing and analysis. However, the visual inspection of individual images is time-consuming and limited by both intra- and inter-rater variance. The difficulty of visual inspection scales with study size and with the heterogeneity of multi-site data. Here, we describe a tool for the automated assessment of Tl-weighted MR images of the brain – MRIQC. MRIQC calculates a set of quality measures from each image and uses them as features in a binary (include/exclude) classifier. The classifier was designed to ensure generalization to new samples acquired in different centers and using different scanning parameters from our training dataset. To achieve that goal, the classifier was trained on the Autism Brain Imaging Data Exchange (ABIDE) dataset (N=1102), acquired at 17 locations with heterogeneous scanning parameters. We selected random forests from a set of models and pre-processing options using nested cross-validation on the ABIDE dataset. We report a performance of ~89% accuracy of the best model evaluated with nested cross-validation. The best performing classifier was then evaluated on a held-out (unseen) dataset, unrelated to ABIDE and labeled by a different expert, yielding ~73% accuracy. The MRIQC software package and the trained classifier are released as an open source project, so that individual researchers and large consortia can readily curate their data regardless the size of their databases. Robust QC is crucial to identify early structured imaging artifacts in ongoing acquisition efforts, and helps detect individual substandard images that may bias downstream analyses.