A multidimensional artefact-reduction approach to increase robustness of first-level fMRI analyses: Censoring vs. interpolating.

BACKGROUND

This manuscript describes a new, multidimensional and data-driven approach to identify outlying datapoints from a first-level fMRI dataset.

NEW METHOD

Using three different indicators of data corruption (the fast variance component of DVARS [Δ%D-var], scan-to-scan total displacement [STS], and each scan's overall explained variance [R]), it identifies outlying datapoints while being balanced using Akaike'c corrected criterion (AIC ) to avoid overcorrection. We then explore the impact of censoring, interpolating, or both, to remove a bad scan's contribution to the final timeseries.

RESULTS AND COMPARISON WITH EXISTING METHODS

Our results (using three real-life datasets and extensive simulations) show that motion-corrupted datapoints as well as non-motion related image artefacts are detected reliably. Using several indicators is shown to be an advantage over existing single-indicator solutions in different settings. As a result of using our algorithm, stronger activation (as detected by both T-value and number of activated voxels) and an increase in the temporal signal-to-noise ratio can be seen. The effects of censoring and interpolation are distinct and complex.

CONCLUSIONS

The multidimensional approach described here is able to identify outlying datapoints in fMRI timeseries, with demonstrable positive effects on several outcome measures. While censoring datapoints may be preferable in many settings, the ultimate choice on which approach to choose may depend on the data at hand. Recommendations are provided for different scenarios.