Automatic alignment of brain anatomy in a standard space is a key step when processing magnetic resonance imaging for group analyses. Such brain registration is prone to failure, and the results are therefore typically reviewed visually to ensure quality. There is however no standard, validated protocol available to perform this visual quality control. We propose here a standardized QC protocol for brain registration, with minimal training overhead and no required knowledge of brain anatomy. We validated the reliability of three-level QC ratings (OK, Maybe, Fail) across different raters. Nine experts each rated N=100 validation images, and reached moderate to good agreement (Kappa from 0.4 to 0.68, average of 0.54±0.08), with the highest agreement for “Fail” images (Dice from 0.67 to 0.93, average of 0.8±0.06). We then recruited volunteers through the Zooniverse crowdsourcing platform, and extracted a consensus panel rating for both the Zooniverse raters (N=41) and the expert raters. The agreement between expert and Zooniverse panels was high (kappa=0.76), demonstrating the feasibility of crowdsourcing QC of brain registration. Our brain registration QC protocol will help standardize QC practices across laboratories, improve the consistency of reporting of QC in publications, and will open the way for QC assessment of large datasets which could be used to train automated QC systems.
A Data-driven Approach for General Visual Quality Control in a Robotic Workcell
