Accurate autocorrelation modeling substantially improves fMRI reliability

Given the recent trend towards validating the neuroimaging statistical methods, we compared the most popular functional magnetic resonance imaging (fMRI) analysis softwares: AFNI, FSL and SPM, with regard to temporal autocorrelation modelling. We used both resting state and task-based fMRI data, altogether 10 datasets containing 780 scans corresponding to different scanning sequences and different subject populations. In analyses of each fMRI scan we considered different assumed experimental designs, as well as different spatial smoothing levels and different detrending options. For data used as null data the use of FSL and SPM resulted in much higher false positive rates than the use of AFNI. On the other hand, due to SPM modelling temporal autocorrelation in the least flexible way, it can introduce negative autocorrelations during pre-whitening for scans with long repetition times. For one dataset we observed a big loss of sensitivity when SPM was used. Interestingly, because pre-whitening in FSL and SPM does not remove a substantial part of the temporal autocorrelation in the noise, we observed a relationship that the lower the assumed experimental design frequency, the more likely it was to observe significant activation. Though temporal autocorrelation modelling in AFNI was not perfect, its performance was much higher than the performance of temporal autocorrelation modelling in FSL and SPM. FSL and SPM could improve their autocorrelation modelling approaches for example adopting a noise model similar to the one used by AFNI.