ICA-AROMA (i.e. ‘ICA-based Automatic Removal Of Motion Artifacts’) concerns a data-driven method to identify and remove motion-related independent components from fMRI data. To that end it exploits a small, but robust set of theoretically motivated features, preventing the need for classifier re-training and therefore providing direct and easy applicability. This beta-version package requires installation of Python2.7 and FSL. Read the provided 'Manual.pdf' for a description on how to run ICA-AROMA.
! NOTE: Previous versions of the ICA-AROMA scripts (v0.1-beta & v0.2-beta) contained a crucial mistake at the denoising stage of the method. Unfortunately this means that the output of these scripts is incorrect! The issue is solved in version v0.3-beta onwards. It concerns the Python scripts uploaded before the 27th of April 2015.
Log report (applied changes from v0.2-beta to v0.3-beta):
-
Correct for incorrect definition of the string of indices of the components to be removed by fsl_regfilt:
changed denIdxStr = np.char.mod('%i',denIdx) to denIdxStr = np.char.mod('%i',(denIdx+1))
-
Now take the maximum of the 'absolute' value of the correlation between the component time-course and set of realignment parameters:
changed maxTC[i,:] = corMatrix.max(axis=1) to corMatrixAbs = np.abs(corMatrix) maxTC[i,:] = corMatrixAbs.max(axis=1)
-
Correct for the fact that the defined frequency-range, used for the high-frequency content feature, in few cases did not include the final Nyquist frequency due to limited numerical precision:
changed step = Ny / FT.shape[0] f = np.arange(step,Ny,step) to f = Ny*(np.array(range(1,FT.shape[0]+1)))/(FT.shape[0])