Deep Scrubbing: A Reconstruction-Based Approach for Automatic and Adaptive fMRI Denoising

Functional magnetic resonance imaging (fMRI) is powerful for studying normal brain function and detecting its dis-orders. However, the notorious problem of fMRI is its signal noise and artifacts that can dramatically diminish the reliability, validity, and diagnostic accuracy of subsequent analyses. Frame scrubbing, one of the well-used featured fMRI denoising methods, aims to detect and remove anomaly data volumes, with the anomaly mainly defined by excessive head motion. Despite its simplicity, head motion-based scrubbing is highly influenced by its threshold, making it challenging to compare results among different studies. We propose a novel, parameter-free, automatic, learning-based scrubbing method for fMRI spatiotemporal anomaly detection. It learns normal fMRI spatiotemporal patterns in a Pre-training Phase, followed by anomaly detection in a Scrubbing Phase through the reconstruction errors. Experiments on big datasets demonstrated our method's effectiveness, as well as the enhanced functional connectivity (FC) and its reliability in the subsequent analysis compared to other well-used scrubbing methods.